Skip to content

algovaultlabs/crypto-quant-signal-mcp

v1.35.0MIT

Composite BUY/SELL/HOLD verdicts for crypto perpetual futures — cross-venue signal interpretation, funding-rate arbitrage scanning, market-regime detection, and an on-chain-anchored track record. Built for AI agents.

Changelog

All notable changes to crypto-quant-signal-mcp are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[1.35.0] - 2026-11-16

Added

  • get_support_resistance_zones (Beta). Returns support and resistance zones for one coin on Binance perpetuals, for one to three timeframes. Read-only. Per timeframe: up to 3 support and 3 resistance zones, ranked outward from price, each with its low and high, distance from price, bars touched in the last 90, tests in the last 30, last touch, and whether it flipped between support and resistance after two closes through it. Also returns volume against its 20-bar average and flags such as no zone above (price above every zone in its history). Use with get_trade_call or get_market_regime for the same coin. $0.02 per call over x402, on POST /x402/get_support_resistance_zones.

Changed

  • get_trade_call on the 30m and 1h timeframes follows the same rule as 3m: _receipts.rule_variant "fade" marks a call that takes the opposite side of the raw reading, and the call's reasons say so.

[1.34.0] - 2026-11-08

Added

  • Saved rules for pre_trade_check. Save your rules once with PUT /api/pre-trade-rules and your API key; GET reads them and DELETE clears them, unbilled. Every check that key makes then applies them, and rules sent with a call still win. A rule from your saved profile shows source: "saved", and the response's saved_rules (applied, updated_at) says whether a profile applied.
  • The docs now show how to make pre_trade_check a step your agent cannot skip: the AlgoVault pre-trade gate for Claude Code and Codex, and recipes for LangChain and Freqtrade ("Gate your orders" on /docs).
  • Receipts. An answered pre_trade_check carries a ledger receipt (id, leaf, verify); a refused call carries none. Once its root is anchored on Base, verify://check/{id} or GET /api/verify-check?id= proves the verdict your agent got was recorded unchanged; until then the lookup reads pending.

Fixed

  • Market regime direction. A swing structure making both higher highs and lower lows now reads MIXED, and the trend direction then follows the directional index (+DI/−DI), so a falling market no longer reads TRENDING_UP. Applies to get_market_regime, scan_market_regimes and pre_trade_check's regime; regime ages in scan_market_regimes restart under the new rule.

[1.33.0] - 2026-11-02

Added

  • pre_trade_check — checks a crypto perpetual futures order against your own rules before you place it. Send coin, side, size and hold time, and optionally leverage, stop, target, account equity, order type and a rules object (for example {"maxLeverage": 10}). Each rule passes or fails with its numbers, then GO, CAUTION or WAIT: a rule of yours that fails returns WAIT, and the built-in checks are listed as default rules you can change or switch off. The response adds the isolated liquidation distance on each venue that publishes its margin tiers and whether it hits before your stop, the round trip in fees for a market or limit entry, the accuracy needed to break even over your hold, funding by venue and the market regime. Pay per call over x402 at $0.02, or use plan calls. Informational and read-only: it places no orders. It is served from the hosted endpoint; a local install refuses it, unbilled, with a pointer there.
  • Sign in from inside Claude. Add AlgoVault as a custom connector and choose Sign in when needed; when Claude asks, click Connect and sign in with your email. On a paid plan, sign in with your subscription email and your plan applies.
  • The published track record now states which calls it leaves out of scoring, in methodology.scoringExclusions on /api/performance-public, performance://signal-performance and get_track_record.

Changed

  • Pay-per-call over x402 now settles only calls that are answered, on both the HTTP route and MCP: a refused or failed call is never charged.
  • get_trade_call receipts can carry _receipts.rule_variant. On the 3m timeframe, "fade" marks a call that takes the opposite side of the raw reading, and the call's reasons say so.
  • A call with an argument the tool doesn't take is refused, unbilled, and the refusal names the valid parameters. The published schemas never allowed extra arguments.
  • A market its venue reports as closed returns VENUE_MARKET_CLOSED before any request is sent to that venue.

Fixed

  • Funding for Binance, Bybit and OKX now comes from each venue's own feed at each contract's funding interval, Bitget is included, and a rate that can't be read is skipped.
  • Track-record tiers follow the venue a call was made on, so a symbol that names different instruments on different venues is tiered per venue.

[1.32.0] - 2026-10-24

Added

  • scan_market_regimes — screens a whole venue by market regime in one call. It returns every coin whose regime matches your filter (TRENDING_UP, TRENDING_DOWN, RANGING, VOLATILE) with confidence, trend strength, ADX, volatility ratio and how long the regime has held. Filters: regime, minimum confidence, asset class, liquidity floor, regime age and a watchlist of up to 200 symbols, plus the same rankBy lenses as scan_trade_calls. Sort by confidence, by the newest regime change or by lens rank; page with limit and offset. Billing is one call per returned row, at least one per request; coins that don't match, and symbols the venue doesn't list, are free. A venue or timeframe it can't serve right now is refused with the alternatives and isn't billed. The screener is served from the hosted endpoint; a local install lists it and refuses the call, unbilled, with a pointer there. Documented on /docs and the tools page, and named in llms.txt.
  • GET /api/plans/public gains free.keyless.monthly_calls (the allowance without a key) and free.telegram (monthly_alerts, daily_alerts — the Telegram bot's own free allowance, counted in delivered alerts). Both are additive.

Changed

  • One free-tier allowance changed: 20 calls a month without a key; a free key gives 200 (up to 100/day).

Fixed

  • On the hosted endpoint, search_knowledge and chat_knowledge now answer from the server's own tool list, so they can name every current tool and its parameters, scan_market_regimes included.
  • A local install now refuses chat_knowledge, unbilled, with a pointer to the hosted endpoint, instead of returning a database error.
  • The desktop-extension API-key field no longer describes a retired free-tier restriction.
  • Each card on the integrations index is a single link again; none nests a link inside another.

[1.31.1] - 2026-10-04

Changed

  • Verdict selection now comes from one declared rule per timeframe and trend direction instead of a global mode setting. Served calls are unchanged.

Fixed

  • The MCP-client setup steps in the README and on the integrations pages now match each client's current release: Claude Desktop's connector path, the Claude Code command (--transport http) and Cursor's config file (~/.cursor/mcp.json). Each install claim on those pages is confirmed against the client's own live source.

[1.31.0] - 2026-09-23

Added

  • plugin.json, mcp.json and .cursor-plugin/plugin.json now ship inside the published npm package. An Agent-Plugins-compatible host — Cursor included — can read the server definition straight from the installed package. The files existed in the repository before this release but were not part of the tarball, so no npm consumer could obtain them.

Changed

  • _algovault.compatible_with no longer names any companion package. It is projected from a machine-verified registry and is [] today, which is the truthful answer: no companion package is published. The README previously printed two package names and stated that downstream tools accept the object directly; both statements are gone. A client that rendered those names should stop.
  • Every colour on every public page now comes from one generated theme region, so the site, the docs and the page templates cannot drift apart.

Fixed

  • Signing in with an email that already holds a paid subscription now returns that subscription's key instead of minting a free one, and the page shown afterwards states the real tier and links the billing portal. A free key cannot resolve at /account, so the portal was unreachable for those accounts.
  • The billing-portal link from /account works again. A form-action directive in the page's Content-Security-Policy was blocking the form that opens it.

[1.30.0] - 2026-09-10

Added

  • claim_url and claim_hint in the _algovault envelope of every tool response that carries an auth block — a direct link to keep the key in use, and a heads-up before the cap. Tools with no auth block carry neither.
  • price_usd_6month on every tier of GET /api/plans/public, so six-month prepay pricing no longer has to be hardcoded by a client.
  • Two routes: a one-click email unsubscribe endpoint and a signature-verified delivery webhook. Plumbing for the email lifecycle; no message is sent by this release.

Changed

  • GET /api/plans/public now returns null for the Enterprise price and call limits instead of a published figure. Enterprise is quoted directly. No key was removed.
  • The webhook documentation no longer publishes per-tier call quotas that were retired in August, and no longer tells you to recreate a subscription after an automatic disable — the subscription re-enables itself.
  • The site carries a sign-up control in the header at mobile widths and a conversion band above the footer; the home page's in-page links and copy button now work below the desktop breakpoint.

Fixed

  • Two upgrade claims on the welcome page that were not true have been removed — one about asset coverage, one about alert limits.

[1.29.0] - 2026-09-04

Added

  • get_track_record — a tool that returns the published track record, for clients that bridge tools but not resources. Defaults to the compact aggregate; include opts into the per-asset, per-venue and recent-call sections. Same data the performance://signal-performance resource serves.
  • GET /api/plans/public — an unauthenticated endpoint returning the current free and paid call allowances, so an integrator no longer has to hard-code them.
  • /.well-known/security.txt and /.well-known/api-catalog are now served; both previously returned 404.
  • A DeepSeek Harness integration guide: https://algovault.com/integrations/deepseek-harness

Changed

  • A trade call on a book that is not currently trading now returns HOLD instead of a directional action, and the reasoning says why and when it resumes. Previously a still book was described as a flat market.
  • get_market_regime no longer reports neutral cross-venue funding when the upstream funding request was refused. A degraded read is now distinguishable from a genuine neutral one.
  • performance://signal-performance returns only the venues currently served, and no longer lists a timeframe the engine does not publish.
  • The recording threshold stated in the methodology block now matches the threshold the engine actually applies.
  • The site declares a brand name to search engines and offers a preferred-sources control on the home page.

Fixed

  • The pricing copy no longer states that HOLD verdicts are unbilled. Quota is counted per call, regardless of verdict.
  • Corrected the DeepSeek Harness install steps, the matching integrations card, and a stale pricing link in one integration tutorial.
  • Repaired the database migration chain so a fresh self-hosted install completes; a missing column previously aborted it partway.

Removed

  • BREAKING — BITMART is no longer an accepted value for the exchange parameter on get_trade_call, get_trade_signal, get_market_regime and scan_trade_calls. Calls passing it now return -32602 invalid_enum_value instead of a named refusal. BitMart is retired and the published enum lists only publicly served venues. If you pinned it, switch to a listed venue. Shipped under a minor for the same reason as the v1.28.0 enum change: the venue had already stopped serving before the schema caught up.

[1.28.2] - 2026-08-26

Added

  • A pairing guide for Binance Agent OS: run AlgoVault and Binance's agent toolkit as two servers in one MCP client, so the verdict and the execution come from different places. OAuth at connect time — no API keys on the machine, no request signing — with trading confined to an isolated sub-account and no withdrawal scope. Guide: https://algovault.com/integrations/binance-agent-os

Fixed

  • The contact form is now challenge-protected against automated abuse, and replies go to a mailbox that is actually monitored.

[1.28.1] - 2026-08-24

Changed

  • Clearer tool descriptions across all seven tools, so an agent picking a tool has more to go on. get_trade_signal now points new integrations at get_trade_call, and scan_trade_calls' exchange parameter says plainly that it takes one venue.

Fixed

  • The LobeHub manifest's tool descriptions were out of step with the ones the server actually serves. All seven are now generated from the same source, and a test keeps them that way.
  • /capabilities and tools/list gave different descriptions for get_trade_signal; the alias steering hint reached only one of them. Both now project from one source.

[1.28.0] - 2026-08-21

Added

  • _algovault.auth on every response: authentication outcome is now reported, not inferred. Absent, malformed, unknown and "could not verify" are distinct states.
  • Documented assetClass (get_trade_call, get_trade_signal) and minLiquidityUsd (scan_trade_calls), projected from the live schema so the docs cannot drift from the server.
  • An error-code reference and the scan_trade_calls response envelope in /docs.

Changed

  • A well-formed but unrecognised API key is now refused. Previously it was served free-tier data, which made a bad key indistinguishable from no key.
  • The published exchange enum lists only publicly served venues; retired and shadow venues no longer appear in the schema.
  • /docs no longer describes a multi-step handshake for the HTTP transport — it has been stateless for some time, and the samples now run as written.

Removed

  • BREAKING — EDGEX and WEEX are no longer accepted values for the exchange parameter on get_trade_call, get_trade_signal and get_market_regime. Calls passing either now return -32602 invalid_enum_value. EDGEX is retired and WEEX was never promoted; the published enum lists only publicly served venues. If you pinned either value, switch to a listed venue. Shipped under a minor because the change went live on the API on 2026-08-19 and neither value was ever documented.
  • The sample activity feed on /verify. The page shows verifiable batch data only.

Fixed

  • A broken distribution badge in the README.
  • GET /signup without a plan returned HTTP 400 while serving the correct plan-picker page; it now returns 200.

[1.27.0] - 2026-08-10

Changed

  • Every verdict counts against your allowance, on both the MCP and HTTP rails.
  • New call ladder: Free 200/month plus 100/day · Starter 10,000/month plus 1,000/day · Pro 100,000/month plus 10,000/day.
  • 6-month prepay replaces annual. Starter is $39.90 ($6.65/mo, 33% off monthly) and Pro is $129 ($21.50/mo, 56% off monthly). Choose it at checkout with ?interval=6month. Annual ($79 / $299) was retired on 2026-08-09; ?interval=year redirects to the 6-month checkout.
  • A daily limit now refuses with the hours until it resets, not a monthly figure.

[1.26.0] - 2026-08-06

Added

  • Annual plans. Prepay a year and save: Starter $79/yr ($6.58/mo, 34% off monthly) and Pro $299/yr ($24.92/mo, 49% off monthly). Choose annual at checkout via ?interval=year; monthly is unchanged and Enterprise remains contact-us.

[1.25.0] - 2026-08-02

Changed

  • Breaking: requires Node 22 or newer. The package now declares engines: >=22. Self-hosters on Node 18 or 20 get an engine warning on install — upgrade Node to continue. The hosted endpoint is unaffected.

Added

  • scan_trade_calls takes an optional minLiquidityUsd floor. It screens the scan universe by notional open interest, or by 24h volume where a venue exposes no bulk open interest. Omit it for no floor.
  • Pay-per-call now settles over Circle Gateway on Optimism, alongside the existing Base rail. Point an x402 client at either advertised rail.
  • Webhook subscribers are told when their endpoint is quarantined, and paid tiers get a longer recovery window than free.

Fixed

  • A paid request whose body was dropped in transit could return a defaults-only result on routes where every field is optional. Every payable route now rejects a dropped body.
  • Rejected paid requests explain why, instead of returning an opaque failure.
  • Payment proofs bind to any advertised rail, not only the first.
  • Funding rates annualize at each contract's own funding period, so cross-venue spreads compare like with like.
  • Free-tier quota exhaustion returns one consistent notice across every surface.
  • get_market_regime now receives the caller's licence, so quota applies correctly.
  • 24h change on Aster reads from the correct upstream field.
  • Webhook status reports when an endpoint last worked, not when it was last probed.

[1.24.1] - 2026-07-25

Internal

  • npm publishing now authenticates via OIDC trusted publishing (no long-lived token). No package-code change.

[1.24.0] - 2026-07-24

Added

  • Three new exchanges live: WhiteBIT, BitMart, and XT — AlgoVault's signal engine now covers 15 perpetual-futures venues. get_trade_call, get_market_regime, and scan_trade_calls accept the new venues via the exchange parameter (enum widened — refresh tools/list).

Internal

  • Structural-feature accrual, PFE-metric integrity, seed/label hardening, and ops. No MCP tool-count change (tools/list = 7).

[1.23.3] - 2026-07-18

Changed

  • Corrected the package description — the perp-venue coverage is now stated qualitatively so it cannot go stale as venues are added.

[1.23.2] - 2026-07-17

Changed

  • The default MCP tool set is 7 tools (trade calls, market scanner, market regime, funding arbitrage, and knowledge search/Q&A). MCP clients should refresh tools/list to update. Additional tools remain available to self-hosters behind opt-in environment flags — see the repository docs.

[1.23.1] - 2026-07-14

Added

  • Dedicated channel hub pages — algovault.com/mcp, /rest-api, /webhooks — each with connect-in-minutes code and a short FAQ.
  • Public tools index at algovault.com/tools + a unified Platform navigation menu across the site.

Internal

  • Funnel, attribution, and analytics-infrastructure hardening. No MCP tool surface change (tools/list = 9).

[1.23.0] - 2026-07-03

Changed

  • scan_funding_arb now scans 7 venues (was 3): Hyperliquid, Binance, Bybit, Gate, KuCoin, Aster, and OKX. Funding is annualized at each venue's own interval, and a per-leg liquidity filter drops thin/untradeable spreads. The tool description changed — MCP clients should refresh tools/list.

Internal

  • Ops, monitoring, and infrastructure hardening. No MCP tool-count change (tools/list = 9).

[1.22.1] - 2026-06-30

Fixed

  • Corrective republish: the npm tarball now carries the v1.22.0 "What's new" + de-enumerated coverage copy (release tag-placement fix — the v1.22.0 tag predated the doc/copy commits). No functional change from 1.22.0.

[1.22.0] - 2026-06-30

Added

  • 12 exchanges live. Added ASTER, BingX, Gate, HTX, KuCoin, MEXC, and Phemex to the verified public track record (5 → 12 promoted venues). Per-coin tools (get_trade_call, get_market_regime) and the per-venue PFE win-rate leaderboard now span all 12 — every call Merkle-anchored on Base L2.

Changed

  • /track-record now renders all 12 venues consistently (analyzing strip, exchange filter, and the cross-venue line) — data-driven, so future venue launches surface automatically with no copy change.
  • Venue coverage across tool/manifest descriptions, the landing page, and docs now reads from a live count, keeping the published number accurate as coverage grows.

[1.21.0] - 2026-06-28

Added

  • scan_trade_calls ranking lenses: new rankBy parameter (9 modes + aliases — funding, gainers/losers, volume, open interest, volatility/ATRP, 24h OI change) with per-row rank values.
  • _receipts field on get_trade_call / get_trade_signal / scan_trade_calls — inline conviction, top factors, live track record, and a verify link.
  • Gemini, Kraken & Alpaca exchange pairings (tutorials + keyless demos) at algovault.com/integrations. (Folds in the v1.20.1 pairings, which were staged but never published while the GitHub account was flagged.)
  • Live track-record leaderboard + public referral program (free email-to-key signup, /referral page).
  • x402 pay-per-call coverage extended to the trade-call and scanner routes.

Security

  • x402 per-route price binding + replay idempotency (closes payment downgrade/bypass).
  • Webhook delivery: egress-IP pinning (IPv6/NAT64 bypass closed) + timestamped HMAC signatures.
  • Invalid Stripe keys now default-deny to the Free tier (no prefix-based escalation).
  • Defense-in-depth: signed-value floor, per-IP quota keying, upstream row-identity assertions.
  • Breaking (webhook subscribers only): the payload signature scheme changed — update your verification per docs/WEBHOOKS.md.

Internal

  • Rate-limit unification, seed orchestration, host/PG rightsizing, attribution, and registry/SEO hardening. No MCP tool-count change (tools/list = 9).

[1.20.1] - 2026-06-05

Added — Exchange pairings: Gemini, Kraken, Alpaca

  • Three new pairing tutorials + runnable keyless demos for autonomous trading agents. Each wires AlgoVault's analysis (composite verdict, market regime, and the scan_trade_calls scanner) into the venue's own MCP execution kit — AlgoVault calls it, the venue executes. Browse at algovault.com/integrations; tutorials in algovault-skills. Testnet/demo only.

Internal

  • Cross-venue rate-limit unification, seed orchestration, grid hardening, and SEO/observability. No public MCP tool surface change (tools/list unchanged at 9).

[1.20.0] - 2026-06-04

Added — scan_trade_calls market scanner

  • New MCP tool scan_trade_calls: scan the top-N perps (1–100, ranked by open interest) on a chosen venue in one call; returns actionable BUY/SELL verdicts with confidence and regime. HOLDs are free — quota counts only non-HOLD results.
  • Webhook subscriptions accept a dynamic top:N asset token (assets: ["top:25"]) that tracks the venue's top perps automatically.

Added — TradFi signal upgrades

  • Session awareness on get_trade_call + get_market_regime: new underlying_session + session_note fields; regime caveats while the underlying market is closed.
  • Cross-venue funding sentiment for TradFi symbols: new metrics.funding_by_venue (all 5 venues, 8h-normalized) on get_market_regime; FIXED_PREIPO funding state with explanatory notes for pre-IPO perps.
  • Structured INSUFFICIENT_CANDLES error with suggested_timeframes replaces the plain-string insufficient-history message.
  • Tool display titles in connector UIs (Composite Trade Call, Market Regime Classifier, …).

Fixed

  • Bitget: volume_24h and oi_change_pct now reflect the queried symbol.

Internal

  • Ops, monitoring, and infrastructure hardening. No breaking changes; existing tool parameter shapes unchanged.

[1.19.1] - 2026-05-31

Docs

  • Ship the v1.19.0 "What's new" (Webhook Delivery Service) to the npm registry README. v1.19.0's README was frozen at v1.18.2 because the release tag pointed at the bare npm version commit (which precedes the doc commits) and publish-npm.yml checks out the tag tree — so the published tarball carried the pre-v1.19.0 README. This patch re-publishes with the tag placed on the final commit.
  • Add a webhook CTA to the README and to the v1.19.0 release announcement: set one up at algovault.com/integrations#webhooks; full docs at docs/WEBHOOKS.md.

[1.19.0] - 2026-05-30

Added — Webhook Delivery Service

  • Hosted outbound webhook delivery: subscribe a URL to receive trade_call and regime_shift push events. HMAC-SHA256 signed, idempotent, retried, quota-gated. Endpoints: POST / GET / DELETE /api/webhooks, plus POST /api/webhooks/:id/test. API key required.
  • GET /api/verify-signal?hash=<signal_hash> resolves the verify_url carried in every payload to that call's public, on-chain-anchored record.

Added — Platform integrations

Internal

  • Monitoring, ops, and adapter hardening. No change to the MCP tool surface (tools/list unchanged).

[1.18.2] - 2026-05-25

Added

  • Two new platform integration examples in github.com/AlgoVaultLabs/algovault-integrations:
    • QuantDinger — reference architecture doc for cross-MCP IDE-mediated orchestration. No transform code; AlgoVault MCP and QuantDinger MCP connect through the user's IDE.
    • Cryptohopper — Signaler webhook with HMAC-sha512 signed payloads. First HMAC-authed transport in the integrations repo.
  • New "Connect Your Trading Platform" section on algovault.com/integrations with tiles for the 5 platform integration examples.

Internal

  • Postgres-CPU autopilot state machine + monitoring wrapper template resolver + canary primitives + build-time landing-page SoT injection + admin dashboard external-only filtering. No public surface delta.

[1.18.1] - 2026-05-23

Fixed

  • Added repository field to package.json so npm provenance attestation cross-check succeeds. Substantive content unchanged from v1.18.0.

[1.18.0] - 2026-05-23

Added — 3-minute timeframe goes live on the public Live Track Record

  • The 3-minute timeframe card and Methodology > Evaluation Windows row are now visible at algovault.com/track-record. The 9th evaluated timeframe is live. (The 3m key shipped to /api/performance-public.byTimeframe in v1.16.0; v1.18.0 completes the dashboard-display chain.)

Added — Platform integration examples

  • New public repo at github.com/AlgoVaultLabs/algovault-integrations ships reference implementations of the Verifiable-Signal v1.0 spec (introduced in v1.17.0). Three platforms in this release:
    • AI4Trade (TypeScript) — REST POST with token-in-body auth.
    • Nautilus Trader (Python) — Subclass of Nautilus's Data abstract base.
    • 3Commas (TypeScript) — Signal Bot webhook via per-bot curry-pattern factory.

[1.17.0] - 2026-05-21

Added

  • Free-tier quota visibility. MCP tool responses surface usage progress at 75% and 90% of the monthly quota. At 100% the tool returns a structured error (TIER_LIMIT_REACHED) with an upgrade link. Existing 100-call monthly quota unchanged.
  • Stripe webhook idempotency. checkout.session.completed events are processed exactly once; UTM tags round-trip through Stripe metadata for paid-conversion attribution.

Distribution

  • AlgoVault listed in TensorBlock's awesome-mcp-servers catalog (5th external discovery channel alongside npm registry, MCP Registry, LobeHub, and Claude Desktop DXT manifest).

[1.16.0] - 2026-05-19

Added

  • 3-minute timeframe added to the public track record. PFE win rate and sample count visible at algovault.com/track-record.
  • Integration walkthroughs at algovault.com/integrations for Claude Desktop, Claude.ai, Cursor, Cline, and Codex CLI.
  • Documentation section at algovault.com/docs covering the search_knowledge and chat_knowledge MCP tools.

Improved

  • Knowledge tools (search_knowledge, chat_knowledge) now index blog posts, YouTube videos, and GitHub Discussions. Bundle refreshes weekly.

[1.15.1] - 2026-05-18

Added

  • Chat usage analytics (admin-gated /admin/chat-analytics dashboard) with weekly Telegram digest.

Fixed

  • Track-record dashboard: LATEST TRADE CALLS panel now matches surrounding card surfaces (visual unification).

Distribution

  • npm registry catches up from v1.14.0 (v1.15.0 was published to GitHub + MCP Registry but skipped npm; v1.15.1 publishes to all surfaces simultaneously).

[1.15.0] - 2026-05-18 — v1.15.0 — search_knowledge + chat_knowledge MCP tools (AV-CHAT-MCP-W1)

Added — two new MCP tools indexed automatically from the v1.14.1 KnowledgeBundle

  • search_knowledge MCP tool + POST /api/search HTTP endpoint. BM25 lexical retrieval over dist/knowledge/latest.json. Free, fast, no LLM call, no quota cost. Self-pitching describe-text intentionally reads as an instruction to the calling agent ("Use this BEFORE attempting any tool call to confirm correct parameter usage"). Backed by wink-bm25-text-search@^3.1.2 (MIT). Field weights: name=3, title=3, description=2, content_markdown=1. Response shape: { query, total_results, results[*], _algovault: { bundle_version, bundle_generated_at } }. Public-shape contract: audits/search-knowledge-shape-snapshot-2026-05-18.json.
  • chat_knowledge MCP tool + POST /api/chat HTTP endpoint. Natural-language Q&A with citations. Backed by Claude Haiku 4.5 (default) or Claude Sonnet 4.6 (opt-in). Locked verbatim 6-rule system prompt; prompt caching enabled (cache_control: ephemeral) for ~90% input-cost discount on the cached system block. Citations carry source_type, source_url, title, excerpt. Quota: Free 10/month, Starter 50/month, Pro 200/month, Enterprise 2000/month — tracked separately from trading-tool quotas in the new chat_usage_monthly Postgres table. Response shape: { question, answer, citations[*], model, _algovault: { bundle_version, bundle_generated_at, quota_remaining } }. Token counts remain operator-internal — never leak to the public response. Public-shape contract: audits/chat-knowledge-shape-snapshot-2026-05-18.json.
  • In-process file watcher. Both tools share a single KnowledgeIndex instance that watches dist/knowledge/latest.json via fs.watchFile (poll every 30s) and atomically rebuilds the BM25 docs on mtime change. Any future tool description, integration tutorial, audit snapshot, or npm version bump flows automatically into the next search/chat response within ≤30s — zero manual refresh, zero hand-curated answers.
  • StubLLMProvider graceful fallback. If ANTHROPIC_API_KEY is unset, server still boots (once-only console.warn at startup); chat_knowledge returns recognizable [STUB] <question> text with citations intact. Live-key provisioning is a post-deploy operator step.
  • New deps. @anthropic-ai/sdk@^0.96.0 (MIT, Anthropic), wink-bm25-text-search@^3.1.2 (MIT, GRAYPE Systems), lru-cache@^10.4.3 (ISC, Isaac Schlueter) added as direct production dependencies.
  • Vitest coverage. 30 new tests across 6 files: tests/unit/{knowledge-index,search-engine,result-cache,llm-provider,chat-engine}.test.ts + tests/integration/knowledge-flow.test.ts. Locks: BM25 build/rebuild semantics, atomic swap, cache TTL + LRU + clear, query/limit/cache invariants, stub fallback, chat context building + citations + cache + truncation, fix-at-generator regex canary (no if question.includes(X) return Y shortcuts), public response shapes match snapshot.

Cache-refresh recommended

MCP clients cache tools/list at session start. To see the new tools:

  • Claude.ai / Claude Desktop — toggle the connector off/on
  • Cursor / Cline — restart the MCP server connection

[1.14.1] - 2026-05-18 — Auto-generated per-release knowledge bundle JSON (KNOWLEDGE-ARTIFACT-W1)

Added — KnowledgeBundle artifact through 3 surfaces

  • /knowledge/latest.json + /knowledge/:slug.json + /knowledge/index.json public HTTP endpoints. Each serves the auto-generated per-release KnowledgeBundle (12-field allow-list — every MCP tool description, response-shape audit snapshot, integration tutorial, package metadata, and README "What's new" slice). Cache-Control: public, max-age=3600. Slug regex ^v\d+\.\d+\.\d+$ guards against path traversal.
  • MCP resource scheme algovault://knowledge/*. New server.resource() registrations for algovault://knowledge/latest + every versioned bundle. MCP clients listing resources via the tools/list extension can discover and read the bundle in-session.
  • GitHub Release asset. Every npm version git-tag push attaches dist/knowledge/algovault-knowledge-v*.json + latest.json + index.json via the new .github/workflows/release-knowledge.yml workflow (softprops/action-gh-release@v2, first-use in this repo).
  • Generator: scripts/build-knowledge-json.mjs (ESM, Node-20-safe, idempotent). Generator-first design — globs over src/tool-descriptions.ts (4 tool entries) + audits/*-shape-snapshot-*.json (response_shapes) + landing/integrations/*.html (integrations) + README.md (whats_new) + package.json (metadata). Adding any new audit snapshot / integration HTML / tool flows automatically into the next bundle. Zero hand-listed knowledge items.
  • Pure formatter: src/lib/knowledge-formatter.ts + lazy-load store: src/lib/knowledge-store.ts. Allow-list semantics — extra keys are silently dropped, so the public shape is guaranteed even if a buggy generator produces extras.
  • Two-sided PII guard (Data Integrity LAW). DENY: build-time regex over the stringified bundle rejects value bindings of outcome_return_pct or outcome_price. REQUIRE: vitest canary asserts at least one response_shapes[*].forbidden_keys array lists "outcome_return_pct" — proves the term exists AS METADATA in the right place, never as a leaked value.
  • Public-shape snapshot: audits/knowledge-shape-snapshot-2026-05-18.json (allowed_keys / forbidden_keys / error_contract / cache_contract / consumers / drift_check_command). Live drift probe: curl -fsS https://api.algovault.com/knowledge/latest.json | jq -e '<has-chain> and (has("outcome_return_pct") | not)'.
  • Vitest canary: tests/unit/knowledge-bundle.test.ts (15 tests). Locks: schema-valid output, byte-idempotent generator, 4-entry tool runtime shape, response_shapes count matches audits/*-shape-snapshot-*.json glob, integrations count matches landing/integrations/*.html glob, two-sided PII guard, empty examples array, formatter rejects empty input.

Changed — MCP tools/list description for get_trade_signal

  • TRADE_CALL_ALIAS_SUFFIX literal rewritten. OLD: " (Alias for \get_trade_call` since v1.10.0; identical behavior. New agents should call `get_trade_call`.)". NEW: " [ALIAS] This tool is an alias of get_trade_call — same behavior, kept for backward compatibility.". Future tool aliases follow the same [ALIAS]` tag pattern.

Cache-refresh recommended

MCP clients cache tools/list at session start. To pick up the refreshed get_trade_signal alias description — Claude.ai / Claude Desktop: toggle connector off+on. Cursor / Cline: restart MCP server connection.

Dockerfile

  • Stage 1 extended with COPY scripts/build-knowledge-json.mjs ./scripts/ + COPY audits/ ./audits/ + COPY landing/integrations/ ./landing/integrations/ + COPY README.md ./README.md + RUN npm run build:knowledge. The Hetzner host re-builds the image post-git pull via docker compose up -d --build, so the generator runs INSIDE the build context — NOT on the GHA runner. dist/knowledge/*.json is inherited into Stage 2 via the existing wholesale COPY --from=builder /app/dist/ ./dist/.

Why a patch (not minor)

  • New endpoints + new MCP resources are additive; no schema mutation on the existing 3 production tools. The alias-suffix copy change is non-breaking (back-compat alias still exists and works identically). Patch bump per SemVer.

[1.14.0] - 2026-05-18 — Framework integrations live: LangChain · LlamaIndex · Microsoft Agent Framework · CrewAI

Added — README cross-links to 4 framework tutorials

  • README.md overhauled from NPM-readme-DRAFT.md SoT. New Framework integrations subsection mirroring the existing exchange-integrations pattern. 4-row table linking the algovault-skills tutorials:
  • Landing-page mirrors at /docs/integrations/{langchain,llamaindex,maf,crewai} — extended scripts/render-integrations.mjs with a FRAMEWORKS array so the same Tailwind navy/mint canonical-chrome wrap that renders the 4 exchange mirrors now produces 4 framework mirrors alongside. Committed static HTML at landing/integrations/{langchain,llamaindex,maf,crewai}.html.
  • Snapshot lines refreshed from live API: PFE WR 90.4% → 90.5%; total calls 86,093 → 96,898; Merkle batches 33 → 38; asset count 720+ → 730+.

Changed — README + landing mirrors only (no runtime change)

  • Pure README + landing-mirror release. Zero tool schema mutation — no parameter changes, no enum changes, no default changes. MCP clients that have already picked up the v1.13.2 tool-description refresh need no further cache invalidation.

Why a minor (not patch)

  • Framework integrations are a NEW public capability surface cross-linked from the npm registry README. Additive, non-breaking. Minor bump per SemVer + per project release ritual (4 marketing artifacts on framework-grade releases).

[1.13.2] - 2026-05-16 — Tool descriptions rewritten for BM25 + regex retrieval ranking

Changed — tools/list description copy (no behavior change)

  • get_trade_call, scan_funding_arb, get_market_regime descriptions rewritten for Anthropic Tool Search (tool_search_tool_regex_20251119 + tool_search_tool_bm25_20251119) which ranks over tools/list name + description + arg-name + arg-description. Each tool's combined-text (tool description + sum of param describe() strings) now contains ≥15 of 20 canonical keyword phrases observed in AI-agent-builder search vocabulary. Length budget tightened: tool descriptions ≤350 chars, param descriptions ≤80 chars.
  • get_trade_signal alias — same canonical description as get_trade_call + back-compat suffix unchanged.
  • get_market_regime no longer claims "for a Hyperliquid perp" — the description now reflects the actual 5-venue coverage (Binance / Bybit / OKX / Bitget / Hyperliquid) via the same enum that get_trade_call uses; default value remains HL (out of scope for this wave).
  • Param describe() strings tightened — verbose shadow-venue prose dropped from exchange describe (5+2 enum still functional; shadow status lives on the canonical mcp://algovault/venues resource).
  • Zero schema mutation. Same enum members, same default values, same Zod constraints. Only the prose payload of tools/list changed.

Added — keyword + brand-voice canary

  • src/tool-descriptions.ts — new pure-data module exporting the description constants + TOP_20_KEYWORDS lock. Hoisted out of src/index.ts so the canary test can import without triggering bottom-of-file startHttp() / startStdio() bootstrap.
  • tests/unit/tool-description-keywords.test.ts (14 cases) — locks: TOP_20 keyword constant shape; ≥15-of-20 keyword coverage per tool; brand-voice forbidden-phrase canary (/intelligence layer|powerful|seamless|robust|cutting-edge|industry-leading|Quant LLM|Wall Street Quant Brain/i); internal-detail forbidden-phrase canary (wave IDs, archetype labels, AOE substrate mechanics); length budget; alias-suffix structure.

Cache-refresh recommended

MCP clients cache tools/list at session start. To pick up the refreshed tool descriptions — Claude.ai / Claude Desktop: toggle connector off+on. Cursor / Cline: restart MCP server connection.

Why this matters

The Arcade benchmark measured 56% regex / 64% BM25 retrieval at 4,027 tools — keyword tightness compounds with catalog growth. The Stacklok comparison shows gateway-side semantic routers reaching 98% retrieval; that's the medium-term direction but keyword retrieval remains the dominant path for the next ~6 months. This rewrite stays inside the brand-voice LAWs (feedback_public_copy_professional_concise.md + feedback_no_internal_details_in_public_copy.md) while packing the substantive search vocabulary AI-agent builders actually type.

[1.13.1] - 2026-05-16 — README republish (NPM-readme-DRAFT.md → README.md)

Changed — npm-surface README only (no code change)

  • README.md fully replaced from NPM-readme-DRAFT.md SoT: canonical hero badge row (Track Record + ERC-8004 Verified Agent), expanded ## 🪪 ERC-8004 Verified Agent on Base mini-section directly under the Live Track Record block, ## What's new in v1.13.0 block + v1.12.0 highlights recap, Architecture diagram now lists 5 CEX + 2 DEX experimental adapters (Aster + edgeX) explicitly, _algovault.version example bumped to 1.13.0. AgentId placeholder <AGENT_ID_PLACEHOLDER> resolved to live 44544 across 5 anchor sites (hero badge href, identity mini-section bullets + blockquote, what's-new bullet, On-Chain Verification list).
  • Snapshot lines refreshed from live API at publish time: PFE WR 90.4% → 90.5%; total calls 86,093 → 92,995; Merkle batches 33 → 36; asset count 720+ → 730+ (in 3 places — Tools coin description, pricing table assets row, Why AlgoVault crypto+TradFi bullet).
  • No src/, dist/, tests/, landing/, or runtime contract changes. Pure README/marketing patch — publishes new tarball so npm registry README reflects v1.13.0's ERC-8004 surface for crawlers that read npm's dist-tags.latest content (npm caches README from the latest publish).

Added — first AlgoVault on-chain presence beyond Merkle anchoring

  • ERC-8004 agent identity adopted on Base mainnet. AgentId 44544 at canonical Identity Registry 0x8004A169FB4a3325136EB29fA0ceB6D2e539a432 (impl getVersion()=2.0.0). Originally minted 2026-04-12 by archived molthunt-launch.ts with non-canonical metadata; ERC-8004-W1 adopted the existing tokenId via setAgentURI(44544, ipfs://bafkreidx3luset63ky5px35fbykcejwgn2ngwtx66tzgq67pvprc4hv5ki) to land canonical-shape registration JSON on-chain. AgentURI now resolves to a spec-compliant ERC-8004 registration file (type/name/description/image/services[]/registrations[]/supportedTrust[]) with an algovault extension namespace (performance pointer, Merkle anchor pointer, verify/track-record URLs).
  • New GET /api/erc-8004-reputation endpoint — public read-only JSON exposing agentId + identity registry + Basescan link + (deferred) attestation aggregator. Path-3 'pending' shape per Plan-Mode Amendment C: score=null, status="pending", attestation_registry=null until ERC-8004-W2 wires attestation (canonical mainnet ValidationRegistry not deployed as of 2026-05-16; Reputation Registry giveFeedback enforces hard-on-chain self-rejection). 5-min in-memory cache. Shape locked by audits/api-erc-8004-reputation-shape-snapshot-2026-05-16.json (6 sections: allowed_keys / forbidden_keys / error_contract / cache_contract / consumers / drift_check_command).
  • landing/index.html + landing/verify.html + README.md Verified-On-Chain surfaces. Hero trust-mark line + verify-page footer row + README ### Verified on-chain subsection, each citing agentId 44544 + the Basescan deep link. data-tr-field="erc8004_agent_id" span hydrated by extended landing/js/track-record-proxy.js (new /api/erc-8004-reputation fetch).

Added — new library + script primitives

  • src/lib/erc8004.ts — Identity Registry ABI + EIP-55-checksummed Base mainnet address constants (IDENTITY_REGISTRY_ADDRESS, REPUTATION_REGISTRY_ADDRESS, VALIDATION_REGISTRY_ADDRESS=null with rationale comment) + CAIP-10 binding + getBaseRpcUrl() + normalizePrivateKey().
  • src/lib/erc8004-registration-json.ts — pure buildRegistrationJson(opts) returning the canonical ERC-8004 + algovault-extension shape. Forbidden-key canary in tests covers top-level + algovault subtree against /outcome_return_pct|outcome_price|phase_e/.
  • src/lib/erc8004-reputation.ts — pure buildErc8004ReputationBody(opts) for the new /api/erc-8004-reputation endpoint. Same forbidden-key canary.
  • src/scripts/register-erc8004-agent.ts — dual-mode (mint OR --update-uri) script. Mint mode: pin v1 → register → pin v2 → setAgentURI (2 txs). Update-uri mode (this wave): single setAgentURI tx, adopts existing agentId with ownership verification. Idempotency via env-pin (~/.config/algovault/erc8004.env) + balanceOf PRE_MINT canary + Transfer-event scan hint. Pinata IPFS pinning.
  • tests/unit/erc8004-registration.test.ts (16 cases) + tests/unit/erc8004-public-shape.test.ts (12 cases) — schema locks + forbidden-key canaries + idempotency subprocess + snapshot-vs-builder parity.

Archived

  • src/scripts/molthunt-launch.ts → archived/scripts/molthunt-launch.ts.deprecated per Plan-Mode Amendment E (D3 — zero external consumers per repo-wide grep). Referenced non-existent agentIdOf(address) ABI fn (canonical IdentityRegistry has only ownerOf(uint256) + balanceOf(address)); used non-canonical {name, description, website, github} metadata shape. archived/scripts/README.md carries deprecation rationale + restore command.

Path 3 (DEFERRED) — attestation pipeline rolling out in ERC-8004-W2

  • Per-resolved-signal attestation flow NOT yet shipped. Path 1 (ValidationRegistry) blocked because canonical mainnet ValidationRegistry not deployed (per erc-8004/erc-8004-contracts/scripts/addresses.ts MAINNET_ADDRESSES.validationRegistry = TBD - need to mine vanity salts; testnet v0.0.1 contract 0x8004Cb1B... squats at the same Base-mainnet address but is risky to use). Path 2 (client-authorized feedback) contradicts zero-friction product positioning. ERC-8004-W2 will re-evaluate once canonical mainnet ValidationRegistry ships.

[1.12.0] - 2026-05-16 — 2-DEX shadow cohort (Aster + edgeX) + tools/list enum widening

Added — new shadow venues callable via explicit exchange arg

  • Aster (ASTER) — BNB-Chain perp DEX, 410 listed perps, ~$2.14B 24h OI per CoinGecko 2026-05-16. REST API is a near-verbatim Binance Futures clone (fapi.asterdex.com/fapi/v1/*). Live trade-call probe: get_trade_call({coin:"BTC", timeframe:"1h", exchange:"ASTER"}) returns Aster's BTCUSDT verdict.
  • edgeX (EDGEX) — L2 zk-rollup perp DEX, 292 contracts, ~$970M OI per AMBCrypto 2026-05-16. REST API at pro.edgex.exchange/api/v1/public/* with {code, data, msg} envelope; numeric contractId ("10000001") with lazy-init lookup table (byCoin/byId cache, 1h TTL); <COIN>USD symbol naming; SNAKE_UPPERCASE klineType (MINUTE_1/HOUR_1/DAY_1); 4-hour funding cadence (annualized × 2190).

Both venues default to status='shadow' per the EXCHANGE-SHADOW-PROMOTE-W1 state machine. NOT in /api/performance-public.byExchange aggregates until promoted (≥80% PFE WR after asset_count × 10 BUY/SELL calls, day-15 minimum window). Visible via:

  • mcp://algovault/venues MCP resource (returns 7 venues: 5 promoted + 2 shadow)
  • /api/performance-shadow public HTTP endpoint (returns per-venue lifecycle metadata)
  • Daily evaluate-venues cron at 06:00 UTC (advances state machine; Telegram alerts on transitions)

Changed

  • MCP tools/list exchange enum widened 5 → 7 venues. Affects both get_trade_call/get_trade_signal (TRADE_CALL_SCHEMA.exchange) and get_market_regime.exchange. Describe-text gains addendum: 'ASTER' = Aster BNB-Chain perp DEX (shadow — experimental), 'EDGEX' = edgeX L2 zk-rollup perp DEX (shadow — experimental).
  • ExchangeId TypeScript union widened 5 → 7 in src/types.ts.
  • getAdapter() dispatch (src/lib/exchange-adapter.ts) gains AsterAdapter + EdgeXAdapter cases.

Deferred (out of scope for this wave)

  • Lighter (zkSync perp DEX) — Plan-Mode probe surfaced HALT-class finding: /candlesticks endpoint returns HTTP/1.1 403 Forbidden from CloudFront (X-Cache: FunctionGeneratedResponse) for unauthenticated callers (reproduces from both Kuala Lumpur and Hetzner-DE PoPs → function-level auth gate, not geo block). Without OHLCV bars, indicator pipeline cannot compute RSI/EMA/Hurst/squeeze. Architect ratified Path A (defer); file LIGHTER-WHEN-CANDLES-W1 follow-up wave to revisit when (a) Lighter's candles endpoint becomes public or (b) we discover an auth scheme via Lighter's GitHub SDK / app.lighter.xyz web-app cookie. See audits/PILOT-ADAPTERS-W1-endpoint-truth.md §2.c.
  • Auto-seeding of ASTER + EDGEX from the daily restricted-universe cron — src/scripts/seed-signals.ts has hardcoded per-venue branches and was out of C1/C2 scope per Bulk-Spec Scope Rule. Tracked as PILOT-ADAPTERS-SEED-LOOP-W2 follow-up. Until then, ASTER + EDGEX are callable via explicit get_trade_call({exchange:'ASTER'|'EDGEX'}) but won't auto-accumulate from the cron.

Cache-refresh notice for MCP clients

Upgrading? MCP clients cache tools/list at session start. Refresh tool list — Claude.ai / Claude Desktop: toggle connector off+on; Cursor / Cline: restart MCP server connection.

The cached 5-venue enum will reject exchange: 'ASTER' / 'EDGEX' until the client re-reads tools/list. After refresh, the new 7-venue enum + shadow-venue describe-text caveat surface.

Plan-Mode artifact

audits/PILOT-ADAPTERS-W1-endpoint-truth.md — 22-row probe matrix across 3 DEX candidates (Aster, edgeX, Lighter); 15-site identifier diff (3 venues × 5 sites: ExchangeId union / TRADE_CALL_SCHEMA / get_market_regime / dispatch / venues seed); 7 nomenclature inline corrections (spec used pre-1.0 fetchCandles/fetchAssetContext/etc.; production is getCandles/getAssetContext per ExchangeAdapter interface); HALT-class Lighter finding with 3 architect-evaluated paths (defer / degraded-shadow / auth-investigate).


[1.11.1] - 2026-05-15 — TradFi symbol aliasing across all 4 CEX adapters

Added

  • Per-CEX TradFi symbol aliasing in all 4 CEX adapters (src/lib/adapters/{binance,bybit,bitget,okx}.ts). get_trade_call({coin: "GOLD", exchange: "BINANCE"}) now resolves to XAUUSDT instead of returning 400 Bad Request from upstream. New TRADFI_ALIASES constant per adapter — alias-discovery derived from live Binance/Bybit/Bitget/OKX exchangeInfo probes (TRADFI-SYMBOL-ALIAS-W1, 2026-05-15). Aliases:
    • Binance: GOLD → XAU, SILVER → XAG, PLATINUM → XPT, PALLADIUM → XPD
    • Bybit: GOLD → XAU, SILVER → XAG
    • Bitget: GOLD → XAU, SILVER → XAG, PLATINUM → XPT, PALLADIUM → XPD
    • OKX: GOLD → XAU, SILVER → XAG, COPPER → XCU, NATGAS → NG, PLATINUM → XPT, PALLADIUM → XPD
    • Symmetric reverse-aliases in fromBinanceSymbol / fromBybitSymbol / etc. so _algovault.coin in the response envelope surfaces the AlgoVault-canonical name (GOLD), NOT the CEX-native (XAU).
  • src/lib/venue-coverage.ts — static per-venue coverage matrix exporting getVenuesSupporting(coin) + isVenueSupportedFor(coin, exchange). HL-only TIER_3 symbols enumerated (24 symbols: ALUMINIUM, BRENTOIL, BX, CORN, DKNG, DXY, EUR, HYUNDAI, JP225, JPY, KIOXIA, KR200, PURRDAT, RIVN, SKHX, SMSN, SOFTBANK, SP500, TTF, URANIUM, URNM, VIX, WHEAT, XYZ100). Partial-coverage TIER_3 symbols enumerated per-venue (14 symbols including HIMS, XLE, COST, GME, NFLX). Plus COVERAGE_PROBED_AT = '2026-05-15' for staleness audits.
  • TradFiSymbolUnsupportedOnVenueError in src/lib/errors.ts — structured error class. Tool handler emits {error_code: 'TRADFI_SYMBOL_UNSUPPORTED_ON_VENUE', message, coin, requested_exchange, suggested_venues, probed_at} instead of a flat {error: 'upstream 400 ...'}. LLM agents pattern-match on error_code and self-retry against suggested_venues[0].
  • Venue-coverage gate wired in getTradeSignal + getMarketRegime (src/tools/get-trade-call.ts, src/tools/get-market-regime.ts). Fires BEFORE the adapter call to avoid hitting upstream with known-unsupported pairs.
  • Plan-Mode artifacts — audits/TRADFI-SYMBOL-ALIAS-W1-endpoint-truth.md (full identifier-diff + collision-fix audit) + audits/TRADFI-SYMBOL-ALIAS-W1-symbol-coverage.csv (248-row per-coin × per-CEX raw probe data).
  • NEW test suite tests/unit/tradfi-symbol-alias.test.ts (35+ cases) — forward + reverse alias resolution per adapter; getVenuesSupporting invariants (HL-only / partial-coverage / all-5 paths); isVenueSupportedFor convenience helper; forbidden-phrase canary carrying over from v1.11.0.

Refactored

  • src/lib/adapters/bybit.ts + src/lib/adapters/bitget.ts — inline coin + 'USDT' (4 sites each) replaced with toBybitSymbol / toBitgetSymbol helpers (mirror existing toBinanceSymbol / toOKXInstId pattern). Reverse-map fromBybitSymbol / fromBitgetSymbol plumbed into getPredictedFundings so the canonical coin name surfaces on every adapter's output.
  • All 4 adapter symbol-helpers (toBinanceSymbol / fromBinanceSymbol / etc.) marked export so the new unit tests can lock them.

Notes — namespace collision caught in Plan Mode

The naive grep-against-symbol-list emitted SP500 → SPX as an alias candidate on all 4 CEXs (SPXUSDT / SPX-USDT-SWAP exists everywhere). Live spot-price probe revealed the collision: SPX on every CEX (and on HL standard perps) is the SPX6900 memecoin (price ≈ $0.40), NOT the S&P 500 index (price ≈ $7400 on HL). The SP500 → SPX alias was DROPPED from all 4 adapter maps BEFORE code-edit landed. SP500 reclassified as HL-only in venue-coverage.ts. Documented in audits/TRADFI-SYMBOL-ALIAS-W1-endpoint-truth.md §3.b. The episode confirms the WIS pattern from v1.11.0: claim-removal and alias-discovery waves need a spot-price (or equivalent semantic-fingerprint) sanity check BEFORE any data table is committed to code.

No client action needed

This is a tool BEHAVIOR change (more permissive resolution), not a schema change. MCP tools/list is unchanged — clients don't need a cache refresh. Existing callers passing exchange: 'HL' for TradFi are unaffected. Callers passing coin: 'GOLD', exchange: 'BINANCE' previously got 400; now get a live verdict via XAUUSDT.


[1.11.0] - 2026-05-15 — Default exchange flipped HL → Binance + README v1.11.0 refresh

Changed (BEHAVIOR — minor bump)

  • Default exchange for get_trade_call and get_trade_signal flipped from 'HL' (Hyperliquid) → 'BINANCE' (Binance USDT-M Futures). Callers omitting the exchange argument now receive the Binance verdict by default. Hyperliquid remains fully supported — pass exchange: 'HL' to preserve prior behavior, or any of the 5 venues explicitly. Rationale: Binance has higher per-IP rate-limit headroom than Hyperliquid (verified live during the wave's Plan-Mode probe — GOLD/HL returned Hyperliquid API rate-limited (429); TSLA/BINANCE + XAU/BINANCE returned live verdicts in the same session), Binance is the most familiar venue for first-time integrators, and the algovault-bot already defaults to BINANCE (BOT-W1 D8 alignment).
  • get_market_regime default exchange is UNCHANGED (still 'HL'). Wave is scoped to call/signal tools per directive.
  • Tool describe-text rewritten to lead with BINANCE as default and drop venue-order bias.
  • Schema-and-handler default coherent — src/tools/get-trade-call.ts fallback input.exchange || 'HL' → input.exchange || 'BINANCE' (dead-code-equivalent post-Zod-default, flipped for consistency + future-proofing).

Removed (FACTUALITY — stale claim deletion)

  • Dropped the "TradFi assets (GOLD, TSLA, etc.) are HL-only" claim from README, tool describe-text, dashboard, asset-tiers description, llms-full.txt, and welcome-page. The claim was verifiably false at two layers:
    • Exchange landscape (May 2026): Binance ships TSLAUSDT (2026-01-28 launch), XAUUSDT + XAGUSDT (TradFi Perpetuals via Nest Exchange / FSRA regulated); Bybit launched 20 US stock perps + XAU/XAG/CL + 3 global ETFs (April 2026); Bitget ships 79+ TradFi instruments at launch (forex, metals, oil, indices, stock perps via Ondo); OKX announced ICE partnership for tokenized NYSE stocks targeting H2 2026.
    • Repo behavior: src/scripts/seed-signals.ts:446 uses SHADOW-SEED-W1 restricted-universe top-N-by-call-count fan-out across all 5 venues; venue-unsupported pairs self-skip via the existing 'Insufficient candle data' error path inside seedExchange. Confirmed empirically via 2026-05-15 postgres GROUP BY (signal_performance.signals): TSLA seeded on HL/BINANCE/BYBIT/OKX/BITGET (260 total), XAU on BINANCE/BYBIT/OKX/BITGET (497 total — notably zero on HL), MSTR on 5 venues (193), NVDA on 4 venues (124), SPX on 5 (192), COIN on 5 (104), AAPL on 4 (103). Only GOLD is HL-only (125) — but that's a symbol-naming artifact: Binance uses XAUUSDT, not GOLDUSDT; XAU coverage on Binance is heavy.
  • Tier-3 asset description rewritten in src/lib/asset-tiers.ts:26 and the dashboard getPerformanceDashboardHtml Tier-3 row in src/index.ts:~1800: 'TradFi perps — stocks, indices, commodities, FX via Hyperliquid' → 'TradFi perps — stocks, indices, commodities, FX (seeded across Binance, Bybit, Bitget, OKX, and Hyperliquid via demand-driven SHADOW-SEED-W1 fan-out)'.

Added

  • NPM-readme-DRAFT.md transplanted to live README.md as canonical README v1.11.0 source (per Mr.1 directive). Includes: hero substrate-frame line "A self-tuning quant ML model with a published track record"; new primary CTA "🤖 Try Free in Telegram"; "Drop-in for every MCP client" matrix (Claude Desktop, Claude Code, Cursor, Cline, Codex, Windsurf, Continue.dev); "What's new in v1.11.0" block with the default-exchange-change announcement; /how-it-works page reference. Architecture diagram updated to put Binance first as default in the exchange-adapter-layer.
  • audits/CHANGE-DEFAULT-EXCHANGE-W1-endpoint-truth.md — Plan-Mode Step 0 artifact: identifier-diff table (36 sites), system-map edge-mutation table, TradFi cross-CEX seed-and-score empirical confirmation (per-coin × per-exchange postgres matrix + live MCP tools/call probes), test-runner probe, schema baseline.
  • NEW canary test tests/unit/default-exchange-binance.test.ts — locks: TRADE_CALL_SCHEMA Zod default = 'BINANCE', describe-text contains "Binance USDT-M Futures (default)", handler fallback uses 'BINANCE', get_market_regime keeps 'HL', no public surface ships the "HL-only TradFi" claim, package.json version = 1.11.0.

Cache-refresh notice for MCP clients

Upgrading? Refresh tool list — Claude.ai / Claude Desktop: toggle connector off+on; Cursor / Cline: restart MCP server connection. MCP clients cache tools/list at session start.

Source citations (web research, 2026-05-13)


[1.10.8] - 2026-05-08 — Telegram bot launch + per-user attribution loop

Added

  • /api/bot/validate-key (server-internal HTTP). New GET endpoint at src/index.ts returns {valid, customer_id, tier} for a Stripe- backed API key. Two-flag firewall via the existing internal-bypass env (BOT_INTERNAL_BYPASS_ENABLED outer + X-AlgoVault-Internal-Key header matching ALGOVAULT_INTERNAL_BYPASS_KEY inner). Consumed exclusively by the public Telegram bot (github.com/AlgoVaultLabs/algovault-bot) over loopback. NOT exposed via the MCP transport — tools/list still returns 3 tools (get_trade_call, scan_funding_arb, get_market_regime).
  • /welcome page deep-link button. Post-Stripe-checkout, the welcome page now shows a "📱 Connect @algovaultofficialbot" button that opens https://t.me/algovaultofficialbot?start=auth_<api_key>. One click binds the user's Telegram chat to their subscription; bot-side tier-aware quota gate honors paid tiers (no 100/mo cap on the bot for Starter/Pro/Enterprise).
  • tier: 'internal' 4th license tier. Recognized via the X-AlgoVault-Internal-Key header (src/lib/license.ts). Quota counter is bypassed (no tick); request_log.license_tier records 'internal' and the new is_bot_internal boolean column for attribution. 10 vitest cases in tests/license-internal-bypass.test.ts.
  • request_log.is_bot_internal BOOLEAN column. Idempotent ALTER TABLE ADD COLUMN IF NOT EXISTS migration in src/lib/analytics.ts for both Postgres and SQLite backends; existing rows default to false.
  • src/lib/bot-auth.ts (NEW). checkBotInternalAuth(headers) helper for server-internal /api/bot/* endpoints. 8 vitest cases.
  • src/lib/welcome-page.ts (NEW). Extracted getWelcomePageHtml from src/index.ts — pure module, importable in tests without triggering app.listen(port). 5 vitest cases for button rendering + URL encoding + DOM ordering.

Changed

  • monitor.js consecutive-fail tracking + 4× exp-backoff on exchange health checks. Suppresses transient-blip alert spam (1-cycle outages) while keeping persistent-failure escalation intact. (Shipped 2026-05-05, included in this release tag.)

Notes

  • Public response shape, MCP tool list, and free-tier behavior unchanged from v1.10.6 / v1.10.7.
  • Companion bot repo: github.com/AlgoVaultLabs/algovault-bot (BOT-W1
    • BOT-W2). Not bundled with this npm package — discovered via the /welcome button or directly at @algovaultofficialbot.
  • BOT_INTERNAL_BYPASS_ENABLED + ALGOVAULT_INTERNAL_BYPASS_KEY env vars are deployment-side; not required for npx-installed stdio usage of this package. Server-side defaults to false so the new endpoints are inert unless explicitly enabled.

[1.10.7] - 2026-05-06 — README hero refresh + track-record image embed

Changed

  • README hero rebuilt around the live track-record image (docs/screenshots/track-record-2026-05-06.png, 2468×528 PNG, 90.0% PFE Win Rate across 74,733 trade calls, 26 Merkle batches anchored on Base L2). New algovault.com + algovault.com/track-record links surfaced prominently. Snapshot KPI line uses <span data-tr-field="…"> tags so algovault.com can live-replace; npm renders the snapshot value as-is.
  • Snapshot baseline refreshed from prior 89.4%+ / 60,000+ / v1.10.3 to current 90.0% / 74,733 / v1.10.7. All snapshot lines marked <!-- SNAPSHOT-LINE --> for future refetch.
  • Documentation-only release. No API, schema, or tool changes. _algovault.version envelope bumps to 1.10.7. Public response shape unchanged. Free-tier behavior unchanged.

[1.10.6] - 2026-05-02 — README link fix (npm view 404s)

Fixed

  • All 28 relative-path links in README.md (Integrations + Skills tables) now point to absolute URLs at https://github.com/AlgoVaultLabs/algovault-skills/blob/main/.... Previous relative paths (docs/integrations/<exchange>.md, examples/<exchange>/demo.mjs, skills/<slug>/SKILL.md) resolved to 404s on both GitHub (the files live in the algovault-skills repo, not this one) and on npmjs.com (npm renders relative links against https://www.npmjs.com/package/<pkg>/<rel>). Display text unchanged — the rendered tables look identical; only the click-target URLs are fixed.
  • No code or API changes; metadata-only patch release.

[1.10.5] - 2026-04-30 — Shadow-seed 1m + 3m + grid re-sizing

Added — Shadow-mode 1m + 3m signal seeding

  • New cron entries for 1m (every minute, top-5 by call-count universe) and 3m (every 3 minutes, top-20 universe) signals. Signals land in the signals table and accrue PFE/MAE outcomes via the existing backfill pipeline. The eval window for both is "12 candles" (12 minutes for 1m, 36 minutes for 3m), matching the established ≤15m → 12 candles rule.
  • Shadow-mode filter on /api/performance-public.byTimeframe: 1m and 3m keys are stripped from the response unless the env flag SHADOW_REVEAL_TIMEFRAMES (comma-list) explicitly reveals them. After 2 weeks of accumulation, Mr.1 reviews the data via the Sunday Telegram digest and flips the env if the timeframe clears the public-track-record bar (PFE WR ≥ 85% AND samples ≥ 3,000).
  • New --restricted-universe N flag on seed-signals.ts + a getRestrictedUniverse(N) resolver that pulls the top-N coins by historical call-count from byAsset aggregation. Used by 1m + 3m crons to bound CPX22 load.
  • New src/scripts/shadow-digest-weekly.ts runs Sunday 00:00 UTC, queries the prior 7 days of 1m/3m signals, posts a Telegram digest with PFE WR, sample size, top/bottom performers per coin, and PASS/FAIL/INSUFFICIENT_DATA verdict per timeframe.
  • New docs/SHADOW_SEED_DECISION_RUNBOOK.md documents the public-flip + rollback recipes for after 2026-05-14 decision day.

Changed — Cross-asset grid re-sized

  • GRID_TIMEFRAMES re-sized from the v1.9.0 logarithmic ladder ['5m', '15m', '1h', '4h'] (24-cell grid) to ['1m', '3m', '5m', '15m', '30m', '1h', '2h'] (42-cell grid). Drops 4h (rank #6 with 2,124 calls) and adds 1m/3m/30m/2h based on empirical call distribution from byAsset.count rankings.
  • New slow-grid circuit breaker: if 3 consecutive grid refreshes each exceed 30s, the breaker opens for 1 hour and the grid temporarily collapses to the v1.9.0 fallback set. Telegram WARNING on every trip.
  • GRID_CONCURRENCY unchanged at 6 (peak ~12 simultaneous HL roundtrips, 76% rate-limit headroom preserved).

Changed — Track-record dashboard copy hygiene

  • Removed three blocks of stale/redundant copy from the public /track-record page: (1) the "9 of 11 timeframes / sub-5m noise-dominated" disambiguation paragraph (it's premature given the empirical test now underway and would be wrong copy if 1m/3m pass), (2) the Methodology "Note" paragraph about directional entry calls (redundant with the PFE Win Rate definition above it), (3) the first two sentences of the bottom italic disclaimer (redundant with the removed Methodology Note); the financial-advice + past-performance sentences are preserved verbatim for legal cover.

Tests

  • tests/unit/shadow-seed-restricted-universe.test.ts (NEW, 7 tests) — resolver + cache + fallback behavior.
  • tests/unit/grid-circuit-breaker.test.ts (NEW, 9 tests) — re-sized grid shape + circuit breaker open/close.
  • tests/unit/shadow-mode-filter.test.ts (NEW, 7 tests) — env-driven filter behavior across 1m/3m + non-shadow timeframes.
  • tests/unit/shadow-digest-format.test.ts (NEW, 6 tests) — digest formatter + PASS/FAIL/INSUFFICIENT_DATA verdict thresholds.
  • tests/cross-asset-grid.test.ts updated — 24-cell expectations re-pegged to 42-cell new shape.
  • tests/unit/cross-asset-grid-backoff.test.ts updated — failure-ratio thresholds re-scaled to 42-cell baseline.

Upgrading

No code changes required. The byTimeframe aggregation stays at 9 keys (5m–1d) until Mr.1 flips the public reveal env. Existing MCP clients calling get_trade_call for 1m or 3m have always worked (the API enum already accepted them); the only change is that those calls will now warm faster on top-5 / top-20 assets via the new grid.

[1.10.4] - 2026-04-30 — README polish

Changed

  • README "What's new in v1.10.3" trimmed: dropped the internal-context "Track-record disambiguation" bullet that exposed the 9-vs-11-timeframes nuance ("sub-5m indicators noise-dominated by design"). That nuance lives in brand-facts.md and on the track-record dashboard where it makes sense in-context; it doesn't belong on the public-facing npm/GitHub README.

No code or API changes — README-only patch.

[1.10.3] - 2026-04-30 — Free-tier unlock + Connect-Your-MCP-Client docs

Changed — Free-tier expansion

  • Free tier now includes all 716 assets and all 11 timeframes (1m, 3m, 5m, 15m, 30m, 1h, 2h, 4h, 8h, 12h, 1d). Was previously gated to BTC + ETH on 15m + 1h only. The 100-calls/month cap remains unchanged as the primary upsell trigger; funding-arb top-5 remains the secondary upsell hook (paid tiers get unlimited funding-arb results). HOLD calls are still always free at every tier.
  • freeGateMessage() reduced to a no-op — coin/timeframe gating is removed; the quota-exhaustion path (getQuotaExhaustedMessage) owns the upgrade- prompt surface from now on. The legacy FREE_COINS / FREE_TIMEFRAMES constants are kept commented-out as reserved emergency-rate-limit-defense switches.
  • Zod schema describe() for get_trade_call's timeframe parameter updated from "Free tier: 15m and 1h only" to "Free tier: all 11 timeframes available, 100 calls/month" so MCP clients render the correct tier capability in their tool-form UIs.

Added — Connect Your MCP Client docs section

  • New <section id="connect-mcp"> rendered into landing/docs.html between <!-- BUILD:mcp-usage:start --> / <!-- BUILD:mcp-usage:end --> markers. Source-of-truth: src/lib/mcp-usage-docs.ts mirroring the signup-flow.ts pattern. Surface / Setup / Result table plus per-client <details> walkthroughs for: Claude Desktop, Cursor, Cline (VSCode), Claude Code, Smithery CLI, plain HTTP / curl. Every config snippet was web-verified against upstream docs on 2026-04-30 with citations + fetch date in the section's footnote so future drift is auditable.
  • scripts/build_landing.mjs extended to handle multiple BUILD blocks (signup-flow + mcp-usage). Idempotent canary preserved (files=0 on second run); --check mode reports drift per-block.

Fixed — Track-record-vs-API-capability disambiguation

  • brand-facts.md §Asset coverage corrected: previously listed 11 timeframes as a forbidden phrase ("currently 9"), but the canonical Zod enum at src/index.ts accepts 11. The 9 visible on the public track-record dashboard are the cron-seeded subset (5m–1d); 1m/3m calls work via API on-demand but don't accrue rolling-window PFE data.
  • Track-record dashboard (/track-record page) gains an explainer line above the "Performance by Timeframe" table clarifying the 9-of-11 distinction. landing/llms-full.txt pricing table gets a coverage disclaimer below it.

Tests

  • tests/unit/license.test.ts (NEW, 62 tests) — free tier accepts every coin + every timeframe; quota / funding-arb caps unchanged.
  • tests/unit/mcp-usage-docs.test.ts (NEW, 14 tests) — MCP_USAGE_HTML structural snapshot.
  • tests/unit/copy-consistency.test.ts (NEW, 113 tests) — grep-guards landing surfaces against legacy free-tier-gating phrases and enforces "11 timeframes" in canonical files.
  • tests/get-trade-signal.test.ts — refreshed: was asserting old "throws with /Starter/" gating behavior, now asserts SOL/1h and BTC/4h succeed on free tier.

Upgrading from v1.9.x or earlier?

MCP clients (Claude Desktop, Claude.ai custom connectors, Cursor, Cline) cache the tool list at session start. The free-tier behavior changed in v1.10.3 — even though no tool was renamed, refresh your tool list so the client picks up the new permissive responses:

  • Claude.ai / Claude Desktop: Settings → Connectors → AlgoVault → toggle off + on (or click "Refresh tools").
  • Cursor / Cline: restart the MCP server connection from the integration panel.

Cached tool responses from before the unlock may still surface "requires Starter" upgrade hints on free-tier calls. Refreshing fixes it instantly. The server is backward-compatible — old code that called get_trade_signal with BTC/ETH on 15m/1h continues to work; it now also accepts every other coin/timeframe combination.

[1.9.0] - 2026-04-15 — Activation patch

Addresses the activation bottleneck identified in experiments/crypto-quant-signal/analytics-funnel-snapshot-2026-04-15.md: 70.5:1 install-to-call ratio, 100% agent one-and-done retention under 1h, 88% of get_trade_signal responses returning HOLD (free tier, zero billable pressure).

Added — Signal surface expansion

  • L2 — HOLD Rescue. On a HOLD verdict, get_trade_signal responses now include closest_tradeable: the highest-confidence non-HOLD cell (BUY or SELL, confidence ≥ 52) from the pre-computed cross-asset grid, excluding the requested (coin, timeframe). Omitted when the grid has no non-HOLD cell. This is a data surface, not a trade recommendation.
  • L4 — Next-Calls Hints. Every response (HOLD and non-HOLD) now includes try_next: the top-3 highest-confidence non-HOLD cells from the same grid, excluding the requested (coin, timeframe). Omitted when the grid is empty. Exposes what the scorer currently sees across the grid; agents decide execution.
  • The grid is computed once per 60-second TTL across 6 assets (BTC, ETH, SOL, BNB, XRP, DOGE) × 4 timeframes (5m, 15m, 1h, 4h) = 24 cells, promise-coalesced, lazy-refresh (no cron, no background worker). See src/lib/cross-asset-grid.ts.

Added — Retention analytics

  • L3 — Session Cohort Surfacing. The mcp-session-id header extracted at the transport layer is now surfaced in every tool response's _algovault.session_id field (null under stdio transport). Per-session cohort metadata is persisted in a new agent_sessions table (first_seen, last_seen, call_count, tools_used, tiers_seen) so retention cohorts are directly queryable without reconstructing from aggregates. Unblocks data-gap #2 from the funnel snapshot.

Changed

  • package.json version bumped 1.8.1 → 1.9.0.
  • Eliminated hardcoded version: '1.x.x' string literals across src/tools/*.ts and src/index.ts. All envelopes and the HTTP /health endpoint now read the version from a shared PKG_VERSION helper at src/lib/pkg-version.ts.

Not in this release

  • L1 — Inline track_record surfacing (re-exposing signal-performance on every get_trade_signal response) is gated on Phase E ✅ (2026-04-17 ~02:30 UTC) and will ship in a follow-up patch.
  • Forum-post cron hardening (Hashnode, Moltbook) shipped on a concurrent branch (harden/agent-forum-post-verify, now merged).

[1.8.1] - 2026-04-13

Changed

  • Website branded around 5-exchange support (Hyperliquid, Binance, Bybit, OKX, Bitget).
  • README rewritten to highlight multi-exchange coverage and adapter architecture.
  • server.json bumped to v1.8.1 for the MCP Registry listing.

Fixed

  • Dashboard now reports consistent Trade Calls and Evaluated columns across all three tables.
  • Exchange tab highlight updates instantly on click.
  • Tier cards respond to exchange/tier filters; removed stale "Tier 1-2 + TradFi" tab.
  • HL TradFi coverage limited to top 20 by OI.
  • backfill-outcomes now passes dex:'xyz' for HL TradFi symbols.

CI

  • Auto-publish release post on version bump deploy.
  • Use grep instead of node for version detection on the VPS (no node on host).
  • Auto-sync landing pages to the Caddy serve path on deploy.

[1.8.0] - 2026-04-11

Added

  • Exchange tabs on the public dashboard.
  • Enriched tier cards with per-exchange breakdowns.

Changed

  • Methodology page updated to reflect the 5-exchange signal pipeline.

Fixed

  • Unicode rendering in public dashboard copy.
  • Monitor treats HTTP 429 as alive; backfill alert threshold raised to 50K.