graphlore: Skill for Claude Code

Exploring a codebase with Graphify

The graphlore MCP server (tools are named graphlore_*; they read the Graphify-built graph) exposes a codebase knowledge graph as tools. Reach for it before grepping or reading many files: it answers structural questions ("what connects to X", "which subsystems exist", "where does this flow run") cheaply, returning just the relevant slice instead of whole files.

If the graphlore_* tools are not available, the MCP server isn't connected — see the project README for the .mcp.json / Claude Desktop config.

The server reads one project — the directory it was started with; there is no per-call project argument, so a question about another repo needs another server. graphlore_build requires the graphify CLI on PATH: graphlore has no extractor of its own.

Preconditions: make sure the graph exists and is fresh

1. graphlore_freshness() — is graph.json stale vs. the current git HEAD?
2. If missing or stale: graphlore_build(".", update=True) (AST-only update needs no API key; mode="deep" adds semantic edges but needs a backend key).

Find code by what it does — graphlore_locate

For a behavioral question ("where is retry/backoff handled?", "how are redirects followed?"), reach for graphlore_locate("<natural-language question>") first: one call runs semantic search, maps the top hit to its enclosing graph node, and returns the token-budgeted subgraph around it plus hidden_links — semantically similar code that is structurally disconnected (duplication / missing-abstraction / sync-async-twin candidates that neither search nor the graph surfaces alone). Works across Python, JS/TS, Go, Java, Rust, C++ and 165+ languages. Needs the optional [semble] extra; without it, fall back to graphlore_search (name-based) and the structural flow below.

Canonical flow (cheap → targeted)

1. graphlore_overview() — ALWAYS first. Size, god nodes, community count, surprise edges, and suggested next steps.
2. graphlore_communities() — the major subsystems. Read these like a table of contents.
3. graphlore_subgraph("<node>", hops=2, budget_tokens=1500) — the workhorse. A token-budgeted BFS slice around a node; this is the cheap way to feed the model just the relevant structure. Start from a god node or a community member.
4. graphlore_query("<natural-language question>", budget=1500) — ask the graph directly (BFS/DFS traversal, no LLM key needed).
5. graphlore_node_details("<node>") — resolve a node to its file:line, type, community, and docstring when you need to jump to source.