generated: '2026-09-19'
method: searched
status: published
source: https://kimetsu.dev/docs/how-kimetsu-works/interfaces/
sources:
- https://kimetsu.dev/docs/how-kimetsu-works/interfaces/
- https://kimetsu.dev/docs/install/
- https://kimetsu.dev/docs/remote/
- https://registry.npmjs.org/kimetsu-ai
- https://crates.io/api/v1/crates/kimetsu-cli
- https://github.com/RodCor/kimetsu
description: 'Kimetsu ships its MCP server inside the kimetsu binary: `kimetsu mcp serve` is a local stdio server
  that hosts (Claude Code, Codex, Cursor, Pi, OpenClaw, Gemini CLI) launch per project, exposing ~28 kimetsu_* tools
  over a per-project SQLite brain. `kimetsu setup --host <agent>` / `kimetsu plugin install <host>` write the host-native
  MCP config (.mcp.json + .claude/settings.json for Claude Code, .codex/ for Codex, openclaw.json, a Pi extension)
  plus the hook wiring. There is NO provider-hosted endpoint: the only HTTP MCP surface is Kimetsu Remote (beta),
  a separate `kimetsu-remote` server binary an operator runs on their own machine (bearer tokens, one brain per repo
  at https://<your-server>/mcp/<repo-id>). The tool list below is transcribed from the Interfaces page; inputSchema
  per tool requires a local `kimetsu mcp serve` + tools/list, which this pass did not run.'
deployment:
  mode: local-stdio
  endpoint: null
  install: npm install -g kimetsu-ai && kimetsu setup --host claude-code
  command: kimetsu mcp serve
  package: https://www.npmjs.com/package/kimetsu-ai
  alt_packages:
  - cargo install kimetsu-cli
  - https://github.com/RodCor/kimetsu/releases (prebuilt lean/embeddings archives for Linux, macOS, Windows)
  auth: none
  verified: searched
  transport: stdio
  hosts_supported:
  - claude-code
  - codex
  - cursor
  - pi
  - openclaw
  - gemini-cli
  self_hosted_remote:
    product: Kimetsu Remote (beta)
    package: https://www.npmjs.com/package/kimetsu-remote
    install: npm install -g kimetsu-remote
    serve: kimetsu-remote serve --addr 0.0.0.0:8787 --data /srv/kimetsu-brains --token <secret> --rate-limit 120
    endpoint_template: https://<your-server>:8787/mcp/<repo-id>
    client_wiring: kimetsu plugin install claude-code --remote https://kimetsu.example.com:8787
    auth: bearer token per teammate (KIMETSU_REMOTE_TOKEN), plain HTTP unless --features tls or a TLS proxy
    unauthenticated_routes:
    - GET /healthz
    - GET /metrics (Prometheus text, aggregate-only)
    rate_limit_default: 120 req/min per token (0 = off); 429 when exceeded
    note: Operator-run, not provider-hosted; recorded here so the remote mode is not mistaken for a public endpoint.
      The docs mark it beta with possible breaking changes before stable.
  registry:
    official_mcp_registry: not listed (registry.modelcontextprotocol.io/v0/servers?search=kimetsu returned 0 on
      2026-09-19)
  note: 'The A2A endpoint https://agents.kimetsu.dev/a2a/sidequest is NOT an MCP server: POST tools/list there returned
    JSON-RPC -32601 Method not found (probed 2026-09-19).'
server:
  name: kimetsu
  transport: stdio
  version: 2.8.0
  version_source: npm kimetsu-ai / crates.io kimetsu-cli, published 2026-09-09
  repository: https://github.com/RodCor/kimetsu
  license: MIT OR Apache-2.0
  response_envelope: 'Every tool returns {"ok": true, "usage": {...}} so the host gets usage guidance with the data.'
tools:
- name: kimetsu_brain_context
  description: Retrieve a context bundle (returns skipped true when nothing relevant, at zero overhead)
  category: retrieval
- name: kimetsu_brain_record
  description: Capture a lesson; runs semantic dedup
  category: memory-write
- name: kimetsu_brain_status
  description: Brain health + memory counts
  category: status
- name: kimetsu_brain_insights
  description: Effectiveness analytics over recent runs
  category: analytics
- name: kimetsu_brain_memory_add
  description: Direct memory create
  category: memory-write
- name: kimetsu_brain_memory_list
  description: Direct memory list
  category: memory-read
- name: kimetsu_brain_memory_top
  description: Top memories
  category: memory-read
- name: kimetsu_brain_memory_search
  description: Memory search
  category: memory-read
- name: kimetsu_brain_memory_proposals
  description: Review pending proposals
  category: curation
- name: kimetsu_brain_memory_accept
  description: Accept a pending proposal
  category: curation
- name: kimetsu_brain_memory_reject
  description: Reject a pending proposal
  category: curation
- name: kimetsu_brain_memory_invalidate
  description: Retire a memory
  category: memory-write
- name: kimetsu_brain_memory_blame
  description: Per-run citation attribution
  category: analytics
- name: kimetsu_brain_memory_conflicts
  description: List ingest conflicts
  category: curation
- name: kimetsu_brain_conflict_resolve
  description: Settle an ingest conflict
  category: curation
- name: kimetsu_brain_prune
  description: List (or invalidate) net-negative memories
  category: curation
- name: kimetsu_brain_model_list
  description: Inspect available embedding models
  category: models
- name: kimetsu_brain_model_set
  description: Switch the embedding model
  category: models
- name: kimetsu_brain_reindex
  description: Re-embed the brain after a model change
  category: models
- name: kimetsu_brain_ingest_repo
  description: Index repo files + manifests
  category: ingest
- name: kimetsu_benchmark_context
  description: Task-aware playbook
  category: benchmark
- name: kimetsu_benchmark_record_outcome
  description: Outcome recording for the task-aware playbook
  category: benchmark
- name: kimetsu_bridge_*
  description: Skill registry discovery, import, export, sync (a family of tools; individual names not enumerated on
    the docs page)
  category: bridge
- name: kimetsu_skills_search
  description: Search the cross-harness skill registry
  category: bridge
- name: kimetsu_brain_cite
  description: Record that a memory materially helped
  category: feedback
- name: kimetsu_brain_answer
  description: Grounded, cited answer composed from memory (local model)
  category: retrieval
- name: kimetsu_brain_config_show
  description: Show the effective configuration (named in the configuration docs)
  category: status
tool_count_stated: ~28
tool_count_named: 27
input_schemas: not captured — require a local tools/list against `kimetsu mcp serve`