graphlens-mcp

A free, MIT-licensed MCP server that gives coding agents (Claude Code, Cursor, and compatible clients) a semantic code graph of your project — symbols, cross-file calls, references, imports and cross-language boundaries.

Instead of reading files top-to-bottom or grepping for names, the agent navigates the structure: who calls this function, what does it depend on, what breaks if I change its signature. It is a thin runtime layer over the graphlens analysis engine: graphlens provides the mechanisms (parsing, stable node identity, resolvers); graphlens-mcp owns the storage, freshness and the agent-facing surface.

📖 Documentation: https://neko1313.github.io/graphlens-mcp/

Status: early. The core navigation works; see Known limitations.

Why

Coding agents discover structure the slow way — grep, glob, read one file at a time — rebuilding call paths by hand before the real work even starts. The motivation is the same as every other code-context tool: stop the agent from grepping. The approach is what sets graphlens apart.

Most tools answer this by building their own model of your code — an ad-hoc graph stitched from heuristics, where every tool maps the codebase a little differently and nothing is authoritative. graphlens takes the opposite bet: it builds on the language's own real analysis engines — rust-analyzer, gopls, the TypeScript compiler, the bundled ty type engine — the LSP-grade tooling the industry already trusts. That yields a stable, real picture of the project (who actually calls what, across files and languages), not a bespoke approximation. And a stable foundation is something you can build on: attach context to the parts of a change that matter, auto-extract semantic clusters, answer impact questions reliably.

That foundation is the graphlens engine — parsing, stable node identity, and the resolvers. graphlens-mcp is a smart, agent-facing layer over it, and — honestly — a worked example of how to use the engine: it persists the graph (so the whole thing isn't held in memory), adds a semantic + clustering layer on top, keeps it fresh as you edit, and exposes it to agents as navigation tools plus a bundled skill. From that example it is growing into a self-sufficient system — one that, measured against the market's giants, aims for stable, reproducible results: better in some places, worse in others, but honest about which (see How it compares).

How it compares

graphlens-mcp ships with a reproducible A/B benchmark (benchmarks/) that drives the same agent against three interchangeable code-context MCP servers — graphlens, semble (semantic search), and codegraph (graph index) — plus a no-tools control that measures how much each server adds over the model's own memory. It runs across real Go / Rust / Python / TypeScript codebases and grades answers deterministically against oracle gold (no LLM judge), stratified into SIMPLE lookups vs HARD impact / cross-file questions, and reports accuracy alongside token / tool-call / dollar cost — because a cheaper arm at equal accuracy wins.

📊 Results (10 repos · 3 models, strong → genuinely weak · ~2,400 graded runs — full breakdown, significance tests and reproduction steps at docs: Benchmarks):

	SIMPLE accuracy	HARD accuracy	HARD tokens (median)	HARD completion
graphlens	0.980 – 1.000	0.899 – 0.921	22.4k – 34.1k	≥ 0.959 on every model
codegraph	0.912 – 0.990	0.655 – 0.939	23.2k – 70.0k	drops to 0.765 on the weakest model
semble	0.647 – 0.961	0.555 – 0.850	21.6k – 74.9k	drops to 0.688 on the weakest model
none (control)	0.366 – 0.681	0.453 – 0.685	0.1k – 0.9k	—

Accuracy alone hides the number that matters to a bill: tokens paid per task. graphlens's HARD-tier token spend stays flat (22k–34k) whether the driving model is strong or weak; codegraph's and semble's balloon past 70k on the weakest model — more than double graphlens's ceiling — for a worse answer, not a better one. graphlens is the only arm that stays clearly ahead of the no-tools control and keeps completion above 0.95 at every model tier: on the weakest model tested (gpt-oss-20b) it holds 0.900 HARD accuracy at roughly half the token cost of codegraph. Pairwise Wilcoxon signed-rank tests (matched by task) confirm the gap is statistically significant on the weaker models, not an artifact of a few outlier tasks — see the notebook for per-model p-values and effect sizes.

Install

Requires Python ≥ 3.13 (a constraint inherited from graphlens).

uv tool install graphlens-mcp      # or: pipx install graphlens-mcp

Python language analysis works out of the box (the ty type engine ships as a dependency). Other languages parse immediately and unlock full cross-file semantics once their toolchain is present (Node for TypeScript, the Go toolchain, etc.); without it that language is reported as degraded rather than blocking init.

Quickstart (two commands)

uv tool install graphlens-mcp        # 1. install
cd your-project && graphlens-mcp init  # 2. index + configure your agent

init detects the project's languages, indexes the code into a local graph, writes the MCP server entry into your agent's config and installs the navigation skill. You do not run serve yourself — your agent launches it from the config. Restart the agent and ask it something like "what breaks if I change the signature of create_order?".

Commands

Useful init flags: --root <dir>, --agent claude_code --agent cursor (repeatable), --no-agent, --no-skills, --db <path>.

The graph lives at <project>/.graphlens/graph.db (SQLite). It is a regenerable cache — safe to delete; reindex rebuilds it. Add .graphlens/ to your VCS ignore (the bundled init flow assumes it is not committed).

Supported languages

graphlens-mcp status reports the actual resolver status per language. When a toolchain is missing, that language is reported as degraded (parsed structure, calls/types not fully resolved) with an install hint — it never blocks init.

Agent tools

Three tools — everything a symbol or file needs comes back as a navigable graph node, not a dead grep line. Each response carries a graph-quality status (ok | degraded) so the agent never mistakes a partial answer for a complete one, plus an indexing flag (true when a background reindex is running, so edges may be temporarily incomplete).

search and relations accept either a symbol name or a node id directly — you don't need to look up a node id first. Both cap their response size (a large hit set is ranked by relevance via a small bundled embedding model, not just truncated) and surface true counts (callers_total, references_total, …) when a list is capped, so the agent sees "15 shown of 22" instead of guessing. If the embedding model can't be fetched (e.g. a first run with no network), search transparently falls back to name/content matching.

Freshness model