ngosangns Knowledge Base

CodeGraph

#technology#codegraph#memory#code-intelligence

Updated: 2026-09-24

Source

Definition

CodeGraph is a code knowledge graph focused on speed and "surgical context". A Rust kernel parses every symbol, call edge, and dependency in the codebase into a graph, so an agent needs only a single codegraph_explore command to get exactly the code it needs instead of crawling file by file with grep/glob/Read. It covers 20 languages and installs from one script with no Node.js required. A file watcher keeps the graph fresh while the agent or the user edits files, so there is no manual reindex step. In the project's own benchmark it cut tool calls by 88%, ran 53% faster, and used 62% fewer tokens. It is MIT licensed. Telemetry is on by default and can be turned off.

Key Metrics

  • Stars: ~71.5k (repo created 01/2026)
  • Forks: ~4.6k
  • License: MIT

Installation and Quick Start

# macOS / Linux — no Node.js required
curl -fsSL https://raw.githubusercontent.com/colbymchenry/codegraph/main/install.sh | sh

# or via npm
npm i -g @colbymchenry/codegraph

# 1. Wire up the agent (Claude Code, Cursor, Codex, opencode, Hermes Agent, Gemini, Antigravity, Kiro, GitHub Copilot)
codegraph install

# 2. Build the graph for a project
cd your-project
codegraph init

codegraph upgrade and codegraph uninstall are available to update or remove everything (including the config written into each agent).

Architecture and Performance

  • Native Rust kernel for 20 languages (TypeScript, JS, Java, Python, Go, C, C++, Rust, C#, Ruby, PHP, Swift, Kotlin, Scala, Dart, R, Lua, Luau — Metal/CUDA go through the C++ path); each language ships only when its graph comes out "byte-for-byte identical" to the reference engine on real repos.
  • Auto-sync: a file watcher using OS-native events (FSEvents/inotify/ReadDirectoryChangesW) keeps the graph up to date whenever the agent or the user edits files — no manual reindex command needed.
  • Scales to the machine: worker pool and cache size follow the real core count (container/cgroup-aware) and measured available RAM — a 2-core/6GB VPS still indexes the Linux kernel (70k files) in under 12 minutes; a fully parallel workstation indexes the Swift compiler repo (27k files) in ~100s, and re-syncing a single file takes ~4s.
  • FTS5 full-text search, impact analysis (callers/callees/blast radius).

Benchmark (self-published, 2026-08)

Measured on headless Claude Opus 4.8, with the codegraph CLI blocked in both arms to prevent cheating, across 7 real OSS repos (VS Code, Excalidraw, Django, Tokio, OkHttp, Gin, Alamofire), median of 4 runs:

  • Average: 88% fewer tool calls, 53% faster, 62% fewer tokens, 44% cheaper, 0 file reads on all 7 repos.
  • Publicly disclosed trade-off: context left in the conversation window after a multi-turn session is ~80% higher than with a grep-and-read agent (because it returns one dense payload instead of many small results that get evicted gradually).

Other Features

  • Anonymous telemetry is on by default (only tool/command/language usage statistics; no code, paths, or queries are sent) — turn it off with codegraph telemetry off or DO_NOT_TRACK=1.
  • A SaaS product, the "CodeGraph platform" (PR-level test/impact prediction), is on a waitlist.
  • Verified releases: built through public GitHub Actions, with npm provenance and attested builds.

Pros

Pro Description
No Node.js required Installs via one script that downloads a prebuilt binary and bundles its own runtime
Smooth auto-sync The graph keeps up with changes in real time, no reindex needed
MIT license No commercial restrictions
Transparent benchmark Publishes both the methodology and the downside (lingering context)

Cons

Con Description
Installs via curl | sh The usual supply-chain risk of any installer like this; consider using npm or inspecting the script before running it
Telemetry on by default You must actively turn it off if you don't want to send statistics
Young repo, very fast star growth Reached ~71.5k stars after being created in 01/2026 — growth far above the norm, so verify independently before fully trusting the self-published benchmark

When to Use

  • You need maximum speed and a graph that stays "fresh" while an agent keeps editing code.
  • You are comfortable with the npm/CLI ecosystem and don't mind installing a native binary.
  • You prefer the MIT license and don't want commercial restrictions like those of GitNexus/Repowise.

References:

See also: Graph Comparison · Source Code Indexing Comparison · Code Review Graph