tokenknows-mcp 0.3.0

TokenKnows MCP server — distill AI coding sessions (Claude Code / Codex / Cursor) into weekly reports, ADRs and a knowledge graph on your self-hosted workbench

TokenKnows MCP Server

MCP server for TokenKnows — a self-hosted engineering knowledge workbench that captures AI coding sessions (Claude Code / Codex / Cursor / VS Code) and distills them into structured documents: weekly reports, tech designs, ADRs, incident reviews, long-form books, agent skills and a knowledge graph, via a 5-stage LLM pipeline. Evidence-linked: every distilled claim links back to source session events.

Prerequisites

This server is the bridge between your MCP host and a self-hosted TokenKnows backend (default http://127.0.0.1:8001). Deploy the backend first — see the main repository. Local-first: your data goes only to the backend you configure.

The distilled documents are best viewed in the TokenKnows web UI (default http://127.0.0.1:5173, code/tokenknows-web in the main repo) — view_url links returned by the tools point there and require a logged-in session. To authenticate the MCP server against a protected backend, create an API token in the web UI under 项目设置 → MCP 接入 (Project Settings → MCP Access) and set it as TOKENKNOWS_API_TOKEN.

Install & run

# Run directly (stdio, for Claude Code / Cowork / Cursor)
uvx tokenknows-mcp

# Or install then run
pip install tokenknows-mcp
tokenknows-mcp

# SSE transport for remote / docker setups
tokenknows-mcp --transport sse --port 8765

Claude Code config example

{
  "mcpServers": {
    "tokenknows": {
      "command": "uvx",
      "args": ["tokenknows-mcp==0.3.0"],
      "env": {
        "TOKENKNOWS_API_BASE": "${TOKENKNOWS_API_BASE:-http://127.0.0.1:8001}",
        "TOKENKNOWS_API_TOKEN": "${TOKENKNOWS_API_TOKEN:-}",
        "TOKENKNOWS_DEFAULT_PROJECT": "${TOKENKNOWS_DEFAULT_PROJECT:-proj-demo-001}",
        "TOKENKNOWS_WEB_BASE": "${TOKENKNOWS_WEB_BASE:-http://127.0.0.1:5173}"
      }
    }
  }
}

All values have working defaults for a default local deployment — zero exports needed. (Claude Code officially supports ${VAR:-default} expansion in .mcp.json; for hosts that don't, write literal values.)

Tip: in Claude Code you can instead install the full plugin (MCP server + slash commands + skills): /plugin marketplace add johnnywuj81/tokenknows → /plugin install tokenknows@tokenknows.

Environment variables

Variable	Default	Description
TOKENKNOWS_API_BASE	http://127.0.0.1:8001	Self-hosted TokenKnows backend URL
TOKENKNOWS_API_TOKEN	—	API token (tkk_...) or JWT bearer; create in web UI → 项目设置 → MCP 接入. Optional for default local deployments
TOKENKNOWS_DEFAULT_PROJECT	proj-demo-001	Default project_id for event submission and distill
TOKENKNOWS_WEB_BASE	http://127.0.0.1:5173	Web UI base used to build view_url links (login required to open them)

Tools

• submit_session_events — persist conversation turns into the knowledge base
• distill_document — trigger the 5-stage pipeline (weekly_report / tech_design / adr / incident / book / agent_skill / knowledge_graph)
• list_assets / get_asset / get_asset_chapters — read distilled output
• search_entity — cross-document knowledge-graph entity search

Plus tokenknows://asset/{id} resources and prompt templates for all 7 document types.

Session watcher (daemon)

The package also ships tokenknows-watcher, a standalone session watcher for Claude Code: it tails ~/.claude/projects/*.jsonl, incrementally uploads conversation turns as events (deduped by content hash, state kept in ~/.tokenknows-watcher.json), so /tokenknows:weekly always has material without manual submit_session_events calls.

# foreground, poll every 30s (default)
uvx --from tokenknows-mcp tokenknows-watcher

# custom interval / one-shot for cron
uvx --from tokenknows-mcp tokenknows-watcher --interval 60
uvx --from tokenknows-mcp tokenknows-watcher --once

It reads the same TOKENKNOWS_API_BASE / TOKENKNOWS_API_TOKEN / TOKENKNOWS_DEFAULT_PROJECT environment variables.

License

MIT — source of truth for this package lives in code/tokenknows-api/mcp_server.