Jonathon's AI Wiki

CodeGraph — Pre-Indexed Code Knowledge Graph for Multi-Agent Coding (colbymchenry)

5 min read

Repo: github.com/colbymchenry/codegraph Stars: 19,133 (at ingest 2026-05-23) Language: TypeScript License: MIT Source: raw/gh-star-colbymchenry-codegraph.md

Pre-indexed code knowledge graph positioned as a shared retrieval layer across Claude Code, Codex, Cursor, OpenCode, and Hermes Agent. Headline pitch: “fewer tokens, fewer tool calls, 100% local.” TypeScript, MIT-licensed, created 2026-01-18, actively pushed (last push 2026-05-23). The pattern is structurally adjacent to this wiki’s own QMD local retrieval — but scoped to code repositories rather than markdown wikis.

Key Takeaways

  • Multi-agent compatibility is the headline thesis — one index serves Claude Code, Codex, Cursor, OpenCode, and Hermes Agent. This treats code-intelligence as a shared infrastructure layer rather than a per-IDE plugin.
  • Local-first — no cloud roundtrip for retrieval. The same architectural pattern as QMD (this wiki’s hybrid BM25+vector+LLM-rerank stack) translated from prose to code.
  • Token-reduction framing matches the broader 2026 context-management thesis — agents waste tokens on Glob + Read + Grep walks when a pre-indexed graph could answer the same question with a focused lookup.
  • TypeScript implementation suggests Node.js / npm distribution; pairs with the MCP server pattern (see essential MCP servers) — a graph backend exposed as MCP tools would slot into any of the named agents without per-agent integration.
  • Multi-agent positioning is a structural bet on the CLI-vs-MCP tradeoff — shared MCP infrastructure beats per-agent CLI tooling as the number of agents grows.
  • Star count 19,133 on a ~4-month-old repo (created 2026-01-18). Anomalous but the MIT license and clear architectural pitch raise the prior compared to opaquer repos — see verification section.
  • The “fewer tokens, fewer tool calls” claim is falsifiable — should be reproducible against a baseline Claude Code session walking the same codebase without the graph loaded.
  • No third-party walkthrough, comparison benchmark, or maintainer talk found at ingest. Verification load is entirely on the repo itself.

Implementation

Tool/Service: colbymchenry/codegraph (TypeScript, MIT, local code knowledge graph) Setup: TypeScript stack — most likely npm install + an index-build step against a target repo, then exposed to agents via MCP or a CLI wrapper. Exact integration paths per agent (Claude Code MCP config vs. Codex CLI vs. Cursor extension) need to be read from the in-repo README. Cost: Free / open-source (MIT — clean redistribution). Integration notes:

  • The interesting test is multi-agent shared state — same graph index serving Claude Code in one window and Codex/Cursor in another. If that round-trips cleanly, this becomes a real infrastructure layer.
  • Comparison to QMD: QMD indexes markdown (wiki articles); CodeGraph presumably indexes ASTs / symbol tables / cross-references. Same architectural shape, orthogonal content. Both local, both hybrid-search-likely.
  • Integration with Claude Code likely via MCP — the essential MCP servers article catalogs the protocol surface this would use.
  • The token-reduction claim is the load-bearing one. Verify by counting tool calls + tokens on a baseline Claude Code session (“describe the architecture of this repo”) with vs. without CodeGraph loaded.
  • For Hermes Agent specifically: CodeGraph would slot as a runtime skill or MCP tool — see essential-mcp-servers for the integration pattern.

Verify before citing

This article inherits the verification caveat from the raw stub. Strict-bar treats high stars + young age as a flag, not a disqualifier.

  • Star count 19,133 on a ~4-month-old repo (created 2026-01-18). Same caveat shape as andrej-karpathy-skills (verified real at 146k stars on similar age). MIT license + clean technical pitch raise the prior compared to opaque-vendor repos, but reproduction is still required.
  • “Fewer tokens, fewer tool calls, 100% local” is falsifiable. Baseline session vs. CodeGraph-loaded session on the same codebase, same question — count tokens and tool calls. The delta either confirms or falsifies the headline.
  • Multi-agent compatibility claim spans 5 tools. Each integration is a separate verification — a repo can ship a working Claude Code path while the Codex / Cursor / OpenCode / Hermes paths are stubs. Check each integration before citing it.
  • Indexing scope unknown. Does it index function signatures? Cross-references? Type usage? Comments? Test files? The retrieval quality depends entirely on what made it into the graph.
  • What would falsify this. If reproduction shows: (i) no measurable token / tool-call reduction vs. baseline, (ii) integrations are aspirational rather than working for ≥2 of the 5 named agents, or (iii) the “100% local” claim conceals a cloud dependency in the indexing step — drop the headline framing and downgrade to “experimental index format.”

Try It

  1. Pull the repo and read its README end-to-end: gh repo clone colbymchenry/codegraph && cd codegraph && cat README.md. Confirm the install path, the index-build step, and which agents have working integrations vs. roadmap items.
  2. Build the index against a small, well-understood repo (something <5k LOC you’ve written yourself). Time the build, measure index size on disk.
  3. Run the same architectural question in Claude Code with and without CodeGraph loaded. Track: tool calls, tokens used, answer quality. This is the falsification test.
  4. If the Claude Code path works, try the second-named agent (Codex or Cursor) — multi-agent shared state is the load-bearing thesis. If only one agent integration actually works, downgrade the wiki article from “multi-agent” to “Claude-Code-with-experimental-other-agent-support.”

Graph View

Backlinks