ContextGraph

Code understanding engine for AI agents. Builds a scope-correct graph of your codebase from tree-sitter ASTs, and answers a question in one call — matched modules, relevant symbols, verbatim source, resolved call edges, and blast radius.

Most tools hand an agent pointers and let it go read files. ContextGraph parses your repository into a graph of symbols and the resolved edges between them, stores it in local SQLite, and serves it over the Model Context Protocol. Ask "how does session revocation work?" and one explore call comes back with the modules that matter, the source of the symbols involved, what calls them, and what breaks if you change them.

Languages. Tree-sitter grammars for Kotlin, Java, TypeScript, TSX, JavaScript, Python, Swift, Objective-C and Go. Symbols carry scope-correct identity — a getName in one class is not confused with a getName in another.

Also ingested. Markdown, PDF, SQL schemas and config files land in the same graph, so a design doc and the code implementing it are connected rather than merely co-located.

Local. SQLite on your disk. No service, no upload. LLM calls happen only if you turn them on.

Quick Start

brew tap erenalpaslan/contextgraph
brew trust erenalpaslan/contextgraph
brew install contextgraph

macOS arm64 and Linux x64. The formula brings its own JDK 17, so no system Java is needed. The middle line is not optional: Homebrew 5.1.15 and later refuse to load a formula from a third-party tap until that tap is trusted.

# Set up and index a project
cd /path/to/your/project
contextgraph init
contextgraph index .

# Ask it things
contextgraph search "authentication"
contextgraph report          # GRAPH_REPORT.md + interactive graph.html

# Serve it to an agent
contextgraph serve-mcp

Prefer to run the jar directly? Every release is on Maven Central, unauthenticated:

curl -fsSLO https://repo1.maven.org/maven2/io/github/erenalpaslan/contextgraph-cli/0.2.0/contextgraph-cli-0.2.0-all.jar
java -jar contextgraph-cli-0.2.0-all.jar --help

Working on the ContextGraph sources themselves? ./gradlew :modules:cli:run --args="index ." runs the same CLI without installing, at the cost of a Gradle startup per command.

MCP Integration

mcp bind prints the block Claude Code needs, so nothing has to be typed by hand:

cd /path/to/your/project
contextgraph mcp bind > .mcp.json

{
  "mcpServers": {
    "contextgraph": {
      "command": "sh",
      "args": ["-c", "cd \"$(git rev-parse --show-toplevel 2>/dev/null || pwd)\" && exec contextgraph serve-mcp"]
    }
  }
}

Commit that file: it names contextgraph on PATH rather than one machine's absolute path, and resolves the project root at launch, so it works for every developer on the team and whichever directory they start Claude Code from.

Continuous Indexing

To keep a repository's graph fresh from CI, install the CLI with the composite action and run ci-reindex — it needs no credentials, since it forces LiteLLM off:

- uses: erenalpaslan/context-graph/action@v0.2.0
- run: contextgraph ci-reindex .

Or hand the whole job over to the reusable workflow, which indexes and uploads the graph as a build artifact, leaving what to do with it — commit it, publish it, discard it — and when to run at all to you:

jobs:
  graph:
    uses: erenalpaslan/context-graph/.github/workflows/reindex.yml@v0.2.0

The primary tool

Tool	Description
contextgraph.explore	Answer a natural-language question in one call. Returns matched modules with descriptions, relevant symbols with verbatim source, their resolved edges (each with confidence and resolution rung), and blast radius. Caps its response at a token budget — top-ranked symbols carry full source, the rest carry signature and location and are marked elided.

explore exists because agents choose badly among many similar tools and chain thin calls expensively. Prefer it for almost every question.

Secondary tools

The other ten remain for when you need one specific thing, and mostly return pointers you then have to follow:

Tool	Description
contextgraph.index_project	Index a directory
contextgraph.search_nodes	Full-text + type-filtered search
contextgraph.get_node	Fetch node with provenance
contextgraph.expand_node	BFS neighborhood expansion
contextgraph.find_path	Shortest path between two nodes
contextgraph.get_evidence	Full provenance chain for a node
contextgraph.impact_analysis	Everything pointing at a node — callers, dependents
contextgraph.related_files	Source files associated with a node
contextgraph.build_context	Ranked context bundle for a task description
contextgraph.generate_report	Generate report and visualization

The server also exposes 6 resources (contextgraph://project, …/graph/nodes, …/graph/edges, …/artifacts, …/reports/summary, …/clusters) and 4 prompts.

Benchmarks

Four sides answered the same 33 questions over the same four repositories, scored against the same expected files — the ones each question's own gold facts cite — by mean reciprocal rank and recall. No LLM is in the loop anywhere: the same corpus and the same questions always produce the same numbers.

• ContextGraph (this project) — this repository's own graph, queried through QueryEngine.buildContext.
• CodeGraph (third-party) — @colbymchenry/codegraph 1.5.0, driven as codegraph explore.
• bash (base-system shell only) — stock grep over a never-indexed checkout. No third-party tool is installed, invoked or assumed: this is the floor a developer with nothing but a shell already has, and the honest thing an index has to earn its cost against.
• ripgrep (third-party) — rg over the same never-indexed checkout, given the same derived tokens. A separate install, kept as the comparator the earlier runs used.

These rows all come from one measurement run. Every side was handed the same raw question text, against the same working copies at the same pinned commits, and scored against the same gold-derived expected files with the same metrics; no side's output was re-ranked, filtered or truncated before scoring. The section this replaces had to warn that its comparator rows came from an earlier run. That caveat is gone because the run it described has been superseded, not because it stopped mattering.

Retrieval, per repository — each repo's own questions, its negative control included:

Repo	Metric	ContextGraph (this project)	CodeGraph (third-party)	bash (base-system shell only)	ripgrep (third-party)
calcom (8)	MRR	0.369	0.125	0.152	0.152
calcom (8)	recall@10	33.8%	3.1%	7.5%	7.5%
excalidraw (9)	MRR	0.556	0.244	0.140	0.140
excalidraw (9)	recall@10	42.6%	22.2%	14.8%	14.8%
gin (8)	MRR	0.719	0.813	0.352	0.352
gin (8)	recall@10	92.7%	83.3%	82.3%	82.3%
keycloak (8)	MRR	0.264	0.056	0.223	0.223
keycloak (8)	recall@10	17.3%	8.8%	39.6%	39.6%

The number after each repo is how many questions it contributes. Bold is the leading column in that row, whoever it is.

ContextGraph (this project) does not lead every row, and the rows it loses are in the table above rather than only in the full report: on gin CodeGraph (third-party) leads on MRR, 0.813 against 0.719; on keycloak bash (base-system shell only) and ripgrep (third-party) lead on recall@10, 39.6% against 17.3%. Plain grep pays nothing for an index and still retrieves more than this project's index does on keycloak.

Pooled over the 29 headline questions, which excludes the four negative-control questions the full report scores separately — so the pooled row and the per-repo rows above are different pools and do not average into each other:

Metric	ContextGraph (this project)	CodeGraph (third-party)	bash (base-system shell only)	ripgrep (third-party)
precision@5	20.0%	12.4%	11.0%	11.0%
precision@10	12.4%	6.2%	8.3%	8.3%
recall@5	38.7%	28.0%	27.0%	27.0%
recall@10	46.0%	28.0%	35.1%	35.1%
MRR	0.502	0.326	0.218	0.218

This project's own progression on excalidraw's nine questions is like-for-like across three published runs: 0.133 MRR before ranking became a function of the query, 0.482 after, 0.556 here, once identifier segments were materialised at index time.

Keycloak is measured this time. The previous run published its retrieval side as unmeasured: IndexIntegrityGate refused it over a single gold-cited file, a META-INF/services/ provider-configuration entry that was missing from the index. The gate was not relaxed and no gold fact was dropped — the indexing gap behind it was fixed. File discovery had been treating org.keycloak.credential.hash.PasswordHashProviderFactory as a secret, because the sensitive-filename heuristic matched words in it; a ServiceLoader registration file is named after the interface it registers and never after its own contents, so those entries are now exempt from that heuristic. Keycloak's index now holds 26 of 26 gold-cited files and the gate passes on its own terms.

An index cannot retrieve a file it never indexed, so coverage of the gold-cited files is published beside the scores rather than after them. ContextGraph (this project) holds 19/19 on calcom, 21/21 on excalidraw, 10/10 on gin, 26/26 on keycloak; CodeGraph (third-party) holds 9/19 on calcom, 17/21 on excalidraw, 10/10 on gin, 22/26 on keycloak. Both text-search sides read the working tree directly and so reach every file by construction.

Ingest — what each index cost to build:

Repo	ContextGraph (this project)	CodeGraph (third-party)
calcom	1m 38s, 268.6 MB	1m 25s, 306.7 MB
excalidraw	22.9s, 32.7 MB	17.8s, 36.5 MB
gin	2.2s, 8.1 MB	899ms, 7.9 MB
keycloak	4m 14s, 1.58 GB	1m 53s, 811.6 MB

Building the index costs this project more than it costs CodeGraph (third-party). On every repo where both indexes were built in this run it is slower, by 1.15× on calcom, 1.29× on excalidraw, 2.40× on gin, 2.24× on keycloak. The claim this replaces — that ingest was roughly an order of magnitude slower than the third-party tool's — rested on an earlier run's figures and no longer holds; the widest gap measured here is gin, at 2.2s against 899ms. On keycloak, the largest index here, it also occupies 1.95× the disk (1.58 GB against 811.6 MB). Both text-search sides have no ingest step at all and pay nothing before the first query — which is the number every row above is being compared against. Earlier work on ingest cost (docs/ingest-cost.md) took Keycloak from 142m 55.6s to 8m 16.4s, 17.3×, by batching SQLite writes; the figure in the table is this run's own measurement on its own corpus and machine, not a further speedup claimed on top of it.

What these numbers do not say. 33 questions over four repositories is a much larger claim than the single repository this section used to report, and it is still not a proof: four repositories are not "code in general"; the questions and the gold facts they are scored against were written by this project, which is the single largest thing a reader should discount for; and each figure is one run rather than a distribution. The two text-search columns come out identical at every k measured here, so the margin over ripgrep (third-party) on this corpus is a margin over grep — ripgrep's own engineering bought nothing a stock shell did not already reach, and the baseline should be read as a plain one rather than a strong one. Nothing here measures answer quality: the metric is which files a tool puts in front of you, not what an agent then does with them, and the agent-level axis is measured separately and never mixed with this one. Ingest was timed once per repo per tool on one machine, with no repeats and no variance. And the rows this project loses — gin's MRR, keycloak's recall@10 — are part of the result, not exceptions to it.

The same run, question by question, as a page you can interrogate: docs/benchmarks/index.html lists every one of the 33 questions with all four sides' scores, filters by repository and category, sorts by any metric, and opens a question to show its expected files against each side's own ranked answer. It is generated from the same result document as everything above and needs no server: open it from a clone with the network off and it works.

Full methodology, the per-question breakdown, the flags both text-search sides were given, every skip and every place this project loses: modules/benchmark/results/four-way/BENCHMARKS.md. Per-signal ablation, the cost of each mechanism, and the reasoning behind every signal that did not ship: docs/retrieval-ranking-ablation.md, docs/identifier-segment-vocabulary.md and docs/ingest-cost.md.

CLI Reference

Command	Description
init	Initialize .contextgraph/ in the current directory
index <path>	Index a directory into the knowledge graph
refresh	Re-parse changed files and bring the local graph in sync with the working tree
watch	Keep the local graph current as files change. Opt-in: requires watcher.enabled=true
ci-reindex [path]	CI-only: reindex from scratch and write the committed baseline. Refuses to run with a TTY attached
search <query>	Search the graph. --limit, --min-confidence, --semantic
node <id>	Show a node with its properties and provenance
expand <id> [--depth N]	BFS neighborhood expansion from a node
path <fromId> <toId>	Shortest path between two nodes
report [--output dir]	Generate GRAPH_REPORT.md and graph.html
describe-modules	Author LLM descriptions for modules lacking one, embed them, and flag descriptions made stale by a changed symbol inventory. --regenerate-stale
export [file.json]	Export full graph snapshot to JSON
serve-mcp	Start the MCP server over stdio
mcp bind	Print the .mcp.json block binding this project to the MCP server
config set <key> <value>	Update configuration

How It Works

FileDiscovery → ArtifactTypeDetector → ExtractorRegistry → ResourceExtractor(s)
                                                (TreeSitter, Markdown, PDF, SQL, Config)
                                                                   ↓
                                     ReferenceResolver → EntityResolver → GraphBuilder
                                                                   ↓
                                                   SqliteStorageAdapter (SQLite + FTS5)
                                                                   ↓
                                   CLI / MCP Server / Report / Visualization / Benchmark

Files are discovered and routed to extractors. Source files are parsed to a tree-sitter AST and walked for declarations, which get scope-correct identities rather than bare names. References that cannot be resolved within a file are held and resolved afterwards through a ladder of increasingly permissive strategies, each edge recording which rung resolved it and with what confidence. Nodes and edges are written to local SQLite in batches — through writeArtifactBatch / upsertNodes / upsertEdges on StorageAdapter — with FTS5 for search. Re-indexing skips unchanged files by checksum.

Modules

Module	Description
core	Domain models, graph schema, StorageAdapter and ResourceExtractor contracts
ingest	File discovery, checksum tracking, pipeline, reference resolution ladder
extractors	Tree-sitter, Markdown, PDF, SQL, config, and LLM-semantic extractors
tree-sitter	Native grammar loading and scope-correct symbol identity for 8 languages
graph	Graph builder, entity resolution, JGraphT algorithms
storage-sqlite	SQLite + FTS5 persistence with Flyway migrations and batch writes
query	Traversal, path finding, evidence retrieval, query-aware relevance ranking
mcp-server	MCP server exposing 11 tools, 6 resources and 4 prompts
cli	Command-line interface (Clikt), watcher, CI reindex
report	GRAPH_REPORT.md generation
visualization	Interactive graph.html export (D3.js)
eval	Grades explore answers against curated questions
benchmark	Four-way retrieval and ingest-cost benchmark: this project, CodeGraph, ripgrep and base-system grep

Configuration

.contextgraph/config.json, created by init:

Key	Default	Purpose
includePatterns	["**/*"]	Glob patterns to include
excludePatterns	build dirs, .git, …	Glob patterns to exclude
maxFileSizeBytes	10 MB	Files larger than this are skipped
ignoreSecrets	true	Skip files that look like secrets
moduleRoots	[]	Explicit module roots, when layout is not inferable
watcher.enabled	false	Opt-in file watcher for watch
watcher.debounceMillis	500	Coalesce bursts of file events
watcher.fallbackIntervalMillis	30000	Full rescan interval, covers dropped watches
litellm.enabled	false	Enable LLM-powered semantic extraction
litellm.baseUrl	http://localhost:4000	LiteLLM proxy endpoint
litellm.model	gpt-4o	Model for semantic extraction
litellm.rateLimitPerMinute	10	Rate limit for LLM calls

Semantic Extraction

Optional. Enables LLM-authored module descriptions, module embeddings for search --semantic, and richer concept extraction, all via a LiteLLM proxy:

litellm --model claude-opus-4-7

contextgraph config set litellm.enabled true
contextgraph config set litellm.base-url http://localhost:4000
contextgraph config set litellm.model claude-opus-4-7
contextgraph describe-modules

Everything else works with LLM calls disabled — that is the default, and CI enforces it.

Building

./gradlew build
./gradlew test

Contributor and agent guidance lives in CLAUDE.md.