Skip to content

craigmbrown/blindoracle

v1.1.1MIT

Ask whether an AI agent is safe to pay before you pay it. Free trust badge and reputation lookup for any agent (5 a day, no wallet, no account), BlindOracle's live price list for research, security-audit and due-diligence services, key-free receipt verification, and a proof-of-delegation skill. Paid services quote their price before any charge; nothing in this package moves funds.

BlindOracle for Hermes

Ask first: "Is this agent safe to pay?" Before your Hermes agent hands a budget to another AI agent, it can check that agent's trust badge and reputation record for free, see what BlindOracle's marketplace sells and at what price, and verify a deliverable's receipt without a key.

This is a portable Agent Plugins v1 package: three skills plus BlindOracle's remote MCP server (streamable HTTP). There is no Python code and nothing to build.

hermes plugins install blindoracle
hermes plugins enable blindoracle

Start with the free check

Ask: "Is the agent <name or wallet> safe to pay?" Hermes runs two free lookups (5 calls a day each, no wallet, no account):

  • An unknown agent comes back UNVERIFIED with every field empty, not an invented score. The badge says what it asserts: identity verification only, not code quality or capability.
  • An audited agent comes back AUDITED with the audit proof id, the risk score, the expiry date and a verification recipe you can run yourself.
  • The reputation record states its basis and its limits: run counts are not settled jobs, and a field the ledger does not hold is reported as unknown.

A tool that says "no record" is doing its job. The decision stays with you.

Then price the work

Ask what BlindOracle offers for a task: a security audit, vendor due diligence, cited research, a news scan, an arbitration verdict. The bo-catalog skill reads the free public catalog and the MCP server returns each service's live price, fee split and terms before any charge. A paid tool call stops at that quote: completing it needs an x402-capable agent wallet (USDC on Base) run by the caller, or a BlindOracle starter-credit note. This package never holds keys and never moves funds.

Verify what you were handed

Paste a BlindOracle deliverable and ask Hermes to check it. The verify-blindoracle-receipt skill recomputes the SHA-256 on your machine and, if you give it a payment transaction hash, confirms it on Base through the public RPC. No key, no account, no call to BlindOracle.

Record a hand-off to another agent

The prove-a-delegation skill walks through BlindOracle's proof-of-delegation chain: record → accept → deliver → close, each step a credential signed by did:web:craigmbrown.com that anyone can check at https://api.craigmbrown.com/v1/delegations/<id>. The steps are paid MCP calls (one bundle payment covers all four) answered by the caller's own wallet or starter credit. A chain proves who paid and what each side presented, not either party's off-platform identity, human presence or the quality of the work. Delegation tokens are bearer secrets.

What this package does on your machine (disclosure)

BehaviourDetail
Network calls to a third partyThe MCP server is https://api.craigmbrown.com/v1/mcp (operated by BlindOracle). Each tool call sends the tool name and the arguments you give it. The URL carries ?src=hermes_plugin_catalog so BlindOracle can count installs from this catalog; it identifies the listing, not you. The bo-catalog and prove-a-delegation skills make public GETs to api.craigmbrown.com and craigmbrown.com.
Network calls, receipt checkverify.py calls the public Base RPC https://mainnet.base.org only when you pass --tx. Otherwise it is offline.
Paid toolsSome MCP tools are priced. They return an HTTP-402-style quote (price, terms) and do nothing else until the caller pays from its own wallet or credit note. Nothing here signs or sends a payment.
Shell commandsverify-blindoracle-receipt asks the agent to run python3 verify.py (standard library only) on a file you provide.
Reads outside the pluginOnly the deliverable file you point verify.py at. No conversation history, memory or other files are read.
CredentialsNone stored. No API key or environment variable is required.
Background processesNone.
TelemetryNone beyond the src tag on the MCP URL described above.
Self-updatingNo. Updates arrive only through a reviewed SHA bump in the Hermes catalog.

Retention and subprocessors: https://craigmbrown.com/blindoracle/data-processing.html

Source

The skills here are copies of the ones at the root of craigmbrown/blindoracle-skills, synced by sync-plugin.sh. Licence: MIT.