codegraph

Blast-radius enforcement for coding agents. codegraph builds an AST call/dependency graph of your repo (tree-sitter, 7 languages) and injects a symbol's callers into the agent's context at the moment it edits the file:

codegraph BLAST RADIUS - includes/checkout.php has 47 caller(s) across 12 file(s).
Audit these before changing signatures/behavior:
  - process_order (includes/orders.php:81)
  - Cart.finalize (includes/cart.php:210)
  ...

Code graphs already exist. They answer questions, which assumes someone asks. An LLM mid-task has usually decided the edit is small and doesn't ask - then breaks callers it never opened. codegraph doesn't wait to be queried; a hook fires on every file edit and puts the dependency list in front of the model, unprompted. Works on any model, since reading a caller list requires no particular capability.

The hooks are warn-only and fail open. A warning never blocks an edit, and any failure in the hook path exits silently. The graph is regenerated from source incrementally, so it can't drift from the code.

Install

Requirements: macOS or Linux, Python 3.10+. Claude Code for the hook integration; the MCP server itself works with any MCP client.

git clone https://github.com/Doublehead/codegraph
cd codegraph
./install.sh

Then restart Claude Code and run /codegraph-reindex inside a project to build its graph. Verify with:

claude mcp get codegraph      # should say: Connected

The installer copies the engine to ~/.local/lib/codegraph, creates a venv with pinned dependencies, runs the full test suite to verify the install, copies the hooks, and registers the MCP server and hooks in Claude Code's config. It's idempotent, backs up every config file it touches, and never clobbers existing servers or hooks. Details in INSTALL.md.

Per-project graph data lives in <project>/.codegraph/ - add that to the project's .gitignore.

Using another MCP client: skip the hook parts and register the stdio server binary directly - ~/.local/lib/codegraph/.venv/bin/codegraph.

Uninstall: install/uninstall.sh. Removes the server, hooks, and command registration; leaves per-project .codegraph/ dirs alone.

Related MCP server: codegraph-mcp

Tools

symbol accepts Class.method to disambiguate. On a large unscoped repo, index returns a scope suggestion and the server instructions direct the agent to apply it; nobody hand-writes config.

Resolution model

A call resolves by its shape, derived from the receiver - never by name alone:

Every edge carries a confidence tier:

• exact - the call shape proves a single target. Trustworthy.

• inferred - a single plausible target, receiver type unverified. Disclosed, never silently exact.

• ambiguous - multiple candidates, edges to all. A real caller is never dropped.

• Calls with no in-repo target are recorded unresolved, not faked.

The bias is recall over precision: an uncertain edge is kept and labelled, never dropped, because a missed caller is the failure this tool exists to prevent. Two invariants hold throughout - no wrong exact edge, no dropped caller - and the regression suite enforces both. Ruby mixins (include/extend/prepend) are part of bare-call lookup; the js/ts/tsx dialects share one resolution namespace.

Framework coupling

Most framework wiring never appears as a call. codegraph extracts it as a separate evidence-gated hook edge tier:

• WordPress: add_action/add_filter/do_action/apply_filters/register_rest_route. hooks(entry_points=true) maps the ajax/REST attack surface with unauthenticated endpoints flagged. All callback forms resolve; a collision resolves to nothing rather than the wrong symbol.

• Django signals: @receiver/.connect paired with .send/.send_robust. Evidence- gated, so a socket's .connect/.send can't fabricate an edge.

• Celery: @task/@shared_task paired with .delay/.apply_async.

• Web routes: Flask/FastAPI decorators, Django path()/re_path(), as entry points.

• React: JSX <Component/> usage becomes a caller edge; HOC-wrapped components register as definitions.

Distinct mechanisms never cross-link. Blind spots that static analysis can't see (interpolated hook names, variable callbacks, closures) resolve to nothing and are disclosed, never faked.

Languages

A grammar that fails to load degrades that one language, not the server. Want another? Open an issue.

Branches and worktrees

The graph always reflects the current working tree. File discovery is git ls-files -co (tracked plus untracked, minus ignored) and every file is read and hashed off disk, not from git objects. So the graph sees whatever branch is checked out plus any uncommitted or unmerged edits sitting on top of it.

Branch switches self-heal. git checkout rewrites the mtime on every file that differs between branches, so the incremental pass re-parses exactly those and skips the rest; the background watcher, SessionStart, and the post-edit hook all trigger it, so the graph is current within a couple seconds of a checkout.

One caveat: there is a single .codegraph/graph.db per project directory, shared across branches. It mirrors what is checked out right now, not all branches at once. In the brief window between a git checkout and the next reindex tick, a query can still reflect the branch you just left; run /codegraph-reindex (or index(force=True)) to make it current immediately.

For true parallel branches, use git worktrees. Each worktree is its own directory and the graph resolution walks up to the nearest .codegraph/, so two worktrees get two independent graphs that never touch each other.

Threat model

codegraph is a local stdio MCP server running single-user at the same OS privilege as the calling agent, which already has file and shell access. There is no privilege boundary here. Root guards, parameter clamps, and size caps exist as footgun protection, not as a security boundary. The bug classes that matter are crashes, hangs, data corruption, and wrong exact edges.

Failure handling

• Every hook failure path exits 0 silently. A malformed scope config indexes everything rather than crashing.

• A corrupt graph DB is quarantined and rebuilt from source, disclosed in the result.

• A crash between file commits and edge resolution leaves a persisted flag that forces re-resolution on the next run, so a half-written graph is never trusted.

• Schema migrations check every column any INSERT uses against old DBs on open.

Tests

python tests/test_codegraph.py     # plain python; also pytest-compatible

62 tests: a 6-language ground truth plus regressions from four adversarial audit rounds. Each confirmed defect became a permanent test with decoys. Run it after any change to the parser, resolver, or hooks.

Known limitations

• Shape-based resolution has a precision ceiling versus full type analysis. The long tail lands in inferred/ambiguous, not in a wrong exact.

• Ruby paren-less calls in expression position (x = helper + 1) are not extracted.

• Member-expression JSX (<Animated.View/>) is skipped.

• JS namespace-import resolution requires the alias to match the filename stem.

• Python re-exports (util.helper defined in util/_impl.py) land unresolved.

• Windows is untested; the hook glue is bash.

License

PolyForm Noncommercial 1.0.0. Free for noncommercial use - individuals, personal projects, education, research, charities, government.

Commercial use - including inside a for-profit company's development workflow - requires a commercial license: ccgraphtheory@shaunoster.com.

Required Notice: Copyright (c) 2026 Shaun Oster (https://github.com/Doublehead)