morpheus-grok-mcp

Grok / Cursor Marketplace MCP that routes OpenAI-compatible inference to Morpheus:

Public (default): https://api.mor.org/api/v1
Custom: your Uplink base https://<host>/v1 (auto-normalized; never invent /api on Uplink)

Name: Morpheus Inference (morpheus-inference)
Maintainer (public): NomadicRogue / ABSGrafx
License: MIT — see LICENSE
Status: v0 implementation (this PR) — stdio MCP + setup/policy skills

GitHub: absgrafx/morpheus-grok-mcp · Python package / CLI: morpheus-mcp (python -m morpheus_mcp)

Accepted product design lives in this repo under docs/ (not a private eng tree).

Install

Marketplace / uvx (preferred)

mcp.json runs:

uvx --from git+https://github.com/absgrafx/morpheus-grok-mcp.git morpheus-mcp

Configure secrets via Plugins → Configure (or vault inject). Values are never committed.

Local development

git clone https://github.com/absgrafx/morpheus-grok-mcp.git
cd morpheus-mcp
uv sync --all-extras
uv run morpheus-mcp
# or: uv run python -m morpheus_mcp

Configure (variable names only)

Variable	Meaning
`MORPHEUS_ENDPOINT_MODE`	`public` \| `custom` (default `public`)
`MORPHEUS_BASE_URL`	Custom Uplink URL when `custom` → normalized to `…/v1`
`MORPHEUS_KEY_MODE`	`shared` \| `per_bot` (default `shared`)
`MORPHEUS_API_KEY`	Shared Bearer key
`MORPHEUS_API_KEY_<SLOT>`	Discrete per-bot secrets (`COS`, `CTO`, `CISO`, `CIO`, `COO`, `CMO`, `CFO`)
`MORPHEUS_API_KEYS_JSON`	JSON map escape hatch slot → key
`MORPHEUS_AGENT_SLOT` / `MORPHEUS_AGENT`	Slot id when `per_bot`
`MORPHEUS_MODEL_DEFAULT`	Exact model id when `morpheus_chat` omits `model`
`MORPHEUS_POLICY_MOE_DEFAULT`	Prefer MoE via `morpheus_chat`
`MORPHEUS_POLICY_CURSOR_EXCEPTION`	Native only if operator names exception
`MORPHEUS_POLICY_FALLBACK_ON_DOWN`	On down → native + `Morpheus didn’t work: <reason>`

Slot ids are lowercase role keys (cos, cto, ciso, cio, coo, cmo, cfo) matching those roles. The JSON-map escape accepts arbitrary slot keys when needed. Master / usage-only keys are refused for chat, embeddings, and list_models.

See plugin.json for titles/descriptions and skills/morpheus-setup/SKILL.md for the guided setup.

Tools (v0)

Tool	Upstream
`morpheus_chat`	`POST {base}/chat/completions`
`morpheus_list_models`	`GET {base}/models`
`morpheus_embed`	`POST {base}/embeddings`

Not shipped: morpheus_usage (deferred).

Skills

skills/morpheus-setup — endpoint/key modes, smoke test, secrets UX
skills/use-morpheus-model — MoE default policy; native only on named exception or down+flag

Public vs custom bases

Mode	Base
public	`https://api.mor.org/api/v1` (forced)
custom	Your host normalized to `https://<host>/v1`

Wrong prefix (/api/v1 on Uplink, or bare /v1 on APIGW) → 404. The connector warns on custom normalize; it does not invent /api for Uplink.

Non-goals (v0)

Not shipping any staff CLI / bridge to Marketplace users
Not embedding Uplink / Lumerin / C-Node
Not re-shipping staff allowlist tools as MCP tools
No secrets in this repo (plugin variables / secure install UI only)
No master key for inference
Not claiming APIGW and Uplink share identical billing//usage semantics

Dogfood note

Crews can sideload this listing with endpoint_mode=custom + per-bot slot secrets to exercise the same MoE path the staff bridge used historically. Host-native tools stay for repo/gh work. Full bridge retirement needs the dogfood checklist in docs/DESIGN.md (including a possible host custom-provider hook as P1 — not a v0 publish blocker).

Design

See docs/DESIGN.md — Conceptual / Logical / Physical, Marketplace fields, and accepted Q1–Q5 locks.

Tests

uv sync --all-extras
uv run pytest

License

CI

Pull requests and pushes to main run unit tests via GitHub Actions (uv sync --extra dev then uv run pytest). No live API calls and no secrets in CI.