kg-mcp

name	kg-mcp
description	This skill should be used when the user asks to expose the project graph as MCP tools, register the graphify MCP server (python -m graphify.serve) in Claude config, says 'graphify mcp 서버', 'MCP로 그래프 쿼리', or invokes /kg-mcp [register|status|unregister]. Configures Claude's .mcp.json so subsequent /kg-query and /kg-orient can call graphify MCP tools instead of loading graph.json directly. Edits user config — requires explicit approval.
trigger	/kg-mcp

/kg-mcp — Register graphify MCP Server in Claude Config

Expose the project's graph as MCP tools so other kg skills can query the graph without loading graph.json into the LLM context. The server is the Python module python -m graphify.serve <graph.json> (or the graphify-mcp console script in graphify v0.8.36+) — not a bare CLI flag; --mcp is not a graphify CLI verb. Target graphifyy>=0.8.24 (recommended 0.8.39); the mcp Python package is required — install the dedicated extra uv tool install "graphifyy[mcp]" (or pip install "graphifyy[mcp]"; graphifyy[all] also bundles it), matching graphify's own runtime error guidance.

Activate When

• User asks "graphify mcp 서버", "MCP로 그래프 쿼리", "register graphify mcp"
• User invokes /kg-mcp [register|status|unregister]
• Before running many /kg-query operations on a large graph (MCP avoids reloading graph.json every time)
• After /kg-update produces a fresh graph and user wants subsequent queries to use it via MCP

Do Not Activate When

• One-off graph query → /kg-query (loads graph.json directly, no MCP setup needed)
• graph.json doesn't exist → /kg-update first
• User wants Obsidian Canvas visualization → /kg-canvas (separate output, not MCP)
• User wants to merge multiple graphs → /kg-merge

Preconditions

• graphify CLI v0.8.x installed (which graphify; uv tool install graphifyy). The graphify-mcp console-script launch form needs >=0.8.36; the python -m graphify.serve form works on any 0.8.x (incl. the 0.8.24 floor).
• Target project has graphify-out/graph.json (verify mtime per ~/.claude/skills/kg/references/architecture.md freshness gate)
• ~/.claude.json or project-local .mcp.json exists and is writable

MCP tools exposed (by python -m graphify.serve)

Per graphify v0.8.x serve.py. The launch command in .mcp.json (set up by register --apply) runs <python> -m graphify.serve <graph.json>, which exposes the tools below over stdio.

Tool	Purpose
query_graph	BFS/DFS search with token budget (replaces in-context graph.json loading)
get_node	Full details for a node by label/ID
get_neighbors	Direct neighbors with edge details
get_community	All nodes in a community by ID
god_nodes	Most connected nodes (anchor concepts)
graph_stats	Summary stats (nodes/edges/communities/confidence breakdown)
shortest_path	Shortest path between two concepts (used by /kg-ingest Confirmation gate, /kg-connect, /kg-postmortem pattern detection)
list_prs / get_pr_impact / triage_prs	PR-dashboard tools (v0.8.x). Need a git/PR context — usually irrelevant to a kg corpus; ignore unless graphing a repo with open PRs.

Transport (v0.8.34+): stdio is the default (one local process per developer). To share one graph across a team, serve over HTTP — python -m graphify.serve <graph.json> --transport http --host 0.0.0.0 --port 8080 --api-key "$SECRET" — and point each .mcp.json at http://<host>:8080/mcp. Set --host 0.0.0.0 and --api-key together; the default 127.0.0.1 bind is loopback-only. Every query_graph call is logged to ~/.cache/graphify-queries.log (disable with GRAPHIFY_QUERY_LOG_DISABLE=1). Note: /kg-mcp register --apply only emits the stdio command/args entry — the HTTP client entry needs a url/http-transport shape, so for HTTP you hand-edit .mcp.json to point at http://<host>:8080/mcp (the server side is real, but register does not wire it up).

Subcommands

/kg-mcp (no subcommand) — defaults to status

Bare invocation /kg-mcp (without register|status|unregister) is treated as /kg-mcp status (read-only inspection). Rationale: status is the only read-only subcommand; defaulting to it lets users inspect current registration with minimal typing. Mutations are always blocked from bare invocation by the existing --scope + --apply Authority gate.

When defaulting to status, prepend this note to the output:

Note: /kg-mcp (no subcommand) maps to status (read-only). Mutations require explicit register --scope <user|project> --apply or unregister --scope <user|project> --apply.

/kg-mcp register --scope <user|project> [--apply]

1. Detect target graph.json (cwd or arg).
2. Print the proposed MCP server entry to add to .mcp.json. For a uv tool install of graphify (recommended), use the self-resolving graphify-mcp console script (v0.8.36+) — its shebang carries the right venv interpreter, so you don't have to locate a python that can import graphify:

   {
     "mcpServers": {
       "graphify-<project-slug>": {
         "command": "graphify-mcp",
         "args": ["<absolute-path-to-graph.json>"]
       }
     }
   }
   

   NOTE (graphify v0.8.x): graphify-mcp launches graphify.serve; there is no bare --mcp CLI verb. The graphify-mcp console script requires graphify >=0.8.36 — on the documented >=0.8.24 floor up to 0.8.35 it does not exist, so on those versions (or any install where graphify-mcp is not on PATH) use the <python> -m graphify.serve form instead. That alternative form {"command": "<python>", "args": ["-m", "graphify.serve", "<graph.json>"]} works only when <python> is the exact interpreter that can import graphify — for a uv tool install that is NOT system python (graphify lives in a uv-managed venv; system python3 -c "import graphify" → ModuleNotFoundError). Find it via the graphify-mcp shebang or uv tool dir, or just prefer the console script. The mcp package is required: uv tool install "graphifyy[mcp]" (or pip install "graphifyy[mcp]"; graphifyy[all] bundles it). Verify with graphify-mcp --help, or <that-python> -c "import mcp.server.stdio" before writing config — NOT import graphify.serve, which succeeds even when mcp is absent (serve.py imports mcp lazily) and then the server dies at runtime with ImportError: mcp not installed. Run: pip install "graphifyy[mcp]".
3. Without --apply: print the proposal and STOP. Do not edit any file. Do not interpret a follow-up user "yes"/"go" as approval. The user must re-issue the command with --apply explicitly added to the same scope. This prevents an LLM-mediated escape hatch around the Authority gate.
4. With --apply: edit .mcp.json (user scope = ~/.claude.json, project scope = <cwd>/.mcp.json).
5. --scope is required, no default. If the user omits --scope, stop and ask which scope; do not assume project or user. User-scope edits affect every Claude session, so silently defaulting either way is wrong.
6. Report which MCP tools are now available; instruct user to restart Claude Code session for MCP server to load.

/kg-mcp status

1. Inspect both ~/.claude.json and <cwd>/.mcp.json for graphify-* MCP server entries
2. Report:
   • Whether graphify-* MCP servers are registered
   • For each, the graph.json path they point to
   • Freshness of each pointed-to graph.json (mtime)
3. Outdated-registration sensor (v0.5.5) — scan project for graphs newer than the registered one:
   • Scan scope: cwd recursively, EXCLUDE these dirs: node_modules/, .git/, __pycache__/, *.egg-info/, build/, dist/, vendor/, vendored/, tmp/, .tmp/, _archive/, _examples/, _fixtures/
   • Priority order when multiple graphs exist:
     1. graphify-out/merged-graph.json at project root (cross-corpus union — most generally useful)
     2. Any */graphify-out/graph.json (per-corpus graphs)
   • If a candidate graph has mtime newer than the currently-registered one by ≥ 5 minutes, emit warning:

     ⚠ Outdated registration:
       Registered: <path> (mtime: <ts>)
       Newer available: <newer-path> (mtime: <ts>) — type: merged | per-corpus
       Suggested: /kg-mcp register --scope <X> --apply (with new path)
     

   • 5-minute grace prevents false positives from filesystem time skew or routine re-runs.
4. Read-only — never edits.

/kg-mcp unregister [--scope user|project] [--apply]

1. Show which graphify-* entries exist in target .mcp.json
2. Without --apply: stop and confirm.
3. With --apply: remove the entry, preserve other MCP servers.

Output Contract

MCP register result: PASS | REFUSED | FAIL  | (status: see status block)
Mode: register | status | unregister

Target config: ~/.claude.json | <cwd>/.mcp.json
Target graph: <absolute path to graph.json>
Graph freshness: FRESH | STALE (<N days>)

Proposed/current MCP servers entry — graphify >=0.8.36 with graphify-mcp on PATH -> command = graphify-mcp (as shown); graphify 0.8.24-0.8.35 or graphify-mcp not on PATH -> command = the interpreter that can import graphify, args = ["-m", "graphify.serve", "<graph.json>"]:
  { "graphify-<slug>": { "command": "graphify-mcp", "args": ["<abs-path-to-graph.json>"] } }

Tools that will be / are exposed:
- query_graph, get_node, get_neighbors, get_community, god_nodes, graph_stats, shortest_path

Action taken:
- Without --apply: nothing written; awaiting approval
- With --apply: <added | removed | already-present | not-present>

Restart required: yes | no
Approval needed: yes (without --apply) | no (status mode)

Confidence: high | medium | low

Caveats:
- <stale graph | other MCP servers in same config | none>

Next command:
- /kg-mcp register --scope <user|project> --apply (after reviewing proposal; --scope required)
- /kg-query "<question>"  (after restart, will auto-use MCP if available)

Exceptions and Escalation

• graphify CLI not found → stop, suggest uv tool install graphifyy (recommended over pip).
• Target graph.json missing → suggest /kg-update first; do not register a non-existent path.
• Target graph.json is stale (≥7d) → warn and ask user; do not block, but Caveats must mention.
• .mcp.json is malformed JSON → stop, do not attempt repair. Report exact parse error and require user fix.
• Without --apply flag → never edit config; only print the proposal. A subsequent user "yes"/"approve"/"go" message is not a substitute for --apply. The user must re-issue the same /kg-mcp command with --apply explicitly present.
• --scope omission → stop and ask. There is no default scope; user-scope and project-scope have very different blast radii.
• Editing user-scope ~/.claude.json affects all sessions → require both --scope user AND --apply explicit per Authority Matrix: user config edits = Human approval.
• Never delete other MCP servers when unregistering — only remove graphify-* entries.

Authority

Per ~/.claude/skills/kg/references/authority-matrix.md:

• .mcp.json edits = Human authority (explicit --apply flag required)
• status mode = LLM (read-only)
• This skill never edits wiki/ or graphify-out/

Quality Gates

Before final answer:

• graphify CLI verified installed
• Target graph.json verified to exist (register/unregister modes)
• Without --apply, no file modified
• With --apply, only graphify-* entries touched (other MCP servers preserved)
• Restart-required note when config changed
• Stale graph noted in Caveats