@mubit-ai/mcp

MCP server for MuBit, exposing the current control-plane memory and MAS surfaces as tools for coding agents.

This package now routes MuBit calls through the canonical JS SDK transport layer instead of maintaining its own raw HTTP client. That keeps endpoint/auth behavior aligned with the SDK while still exposing current MuBit-only tools.

Install

npm install -g @mubit-ai/mcp

Or run it directly:

npx @mubit-ai/mcp

Env vars

| Variable | Required | Default | Purpose | | --- | --- | --- | --- | | MUBIT_API_KEY | Yes | — | MuBit instance API key | | MUBIT_ENDPOINT | No | http://127.0.0.1:3000 | MuBit HTTP endpoint | | MUBIT_DEFAULT_SESSION_ID | No | default | Default run/session for tool calls | | MUBIT_DEFAULT_USER_ID | No | — | Default user scope |

MCP setup

Claude Code:

claude mcp add mubit-memory -- npx @mubit-ai/mcp

Generic mcp.json:

{
  "mcpServers": {
    "mubit-memory": {
      "command": "npx",
      "args": ["@mubit-ai/mcp"],
      "env": {
        "MUBIT_API_KEY": "mbt_...",
        "MUBIT_ENDPOINT": "http://127.0.0.1:3000",
        "MUBIT_DEFAULT_SESSION_ID": "coding-session"
      }
    }
  }
}

Tool surface

Core memory tools:

• mubit_remember
• mubit_recall
• mubit_context
• mubit_archive
• mubit_dereference
• mubit_reflect
• mubit_lessons
• mubit_forget
• mubit_status

MAS and learning-loop tools:

• mubit_checkpoint
• mubit_outcome
• mubit_step_outcome
• mubit_learned
• mubit_strategies
• mubit_register_agent
• mubit_list_agents
• mubit_handoff
• mubit_feedback
• mubit_ingest_status

Observability tools:

• mubit_memory_health
• mubit_diagnose

Current parameter coverage

mubit_remember supports current MuBit structured ingest fields:

• intent: fact, observation, lesson, rule, handoff, feedback, tool_output, tool_input, trace, reflection, task_result, log, context
• lesson_type, lesson_scope, lesson_importance
• lesson_conditions
• source
• upsert_key
• metadata

mubit_recall supports:

• entry_types
• include_working_memory
• agent_id
• session_id / user_id

Each recalled evidence item carries a reference_id. Feed those back to mubit_outcome via entry_ids to credit the exact entries that helped.

mubit_outcome records reinforcement and threads the attribution loop:

• reference_id (required; use "global" for run-level reinforcement)
• entry_ids: reference_ids of recalled evidence that contributed
• verified_in_production: applies a retrieval boost

mubit_context supports current context assembly controls:

• mode: full, summary, sections
• sections
• entry_types
• max_token_budget
• agent_id

mubit_archive and mubit_dereference support exact-reference reuse:

• artifact_kind
• origin_agent_id
• source_attempt_id
• source_tool
• labels
• family
• reference_id

Example flows

Checkpoint + reflection:

1. mubit_checkpoint
2. mubit_reflect
3. mubit_lessons

MAS coordination:

1. mubit_register_agent
2. mubit_handoff
3. mubit_feedback
4. mubit_context (sections=["handoffs","feedback"])

Failure diagnosis:

1. mubit_diagnose
2. mubit_memory_health
3. mubit_recall (entry_types=["lesson","rule"])

Attribution loop (credit recalled evidence):

1. mubit_recall (capture evidence[].reference_id)
2. ... act on the recalled guidance ...
3. mubit_outcome (reference_id="global", entry_ids=[<recalled reference_id>...], verified_in_production=true)

Development

cd integrations/mcp
npm install
npm test
npm run lint
npm run build

The published runtime entrypoint is dist/index.js, so dist/ must be regenerated whenever src/ changes.

License

Apache-2.0