Pi Codegraph — Development skill for Claude Code

pi-codegraph

**A trusted Pi-side wrapper around [codebase-memory-mcp](https://github.com/DeusData/codebase-memory-mcp).** It gives an agent *derived* knowledge — call graphs, blast radius, architecture — extracted from your code, so it stops re-greping the same repo every session. The companion to [pi-okf](https://github.com/renezander030/pi-okf) (which carries *authored* knowledge) and a sibling of [pi-gate](https://github.com/renezander030/pi-gate).

codebase-memory-mcp parses a repo into a persistent knowledge graph (tree-sitter, 158 languages) and exposes ~14 MCP tools. `pi-codegraph` is the deterministic, agent-first front end: it resolves and speaks MCP to that binary, normalizes the results to JSON (or a `-H` table), and puts a harness-owned trust boundary in front of it.

`pi-recall` is the standalone Pi extension that can use `pi-codegraph` as its first recall backend. Keep the boundary clear: `pi-codegraph` owns the trusted derived-codegraph wrapper; `pi-recall` owns the Pi slash commands and model-callable recall tool.

Why a wrapper

Some knowledge an agent needs is already in the code: who calls a function, what a diff breaks, where the dead code is. No graph tool should be trusted blindly with your repo, and an agent should not be re-launching a third-party server with ad-hoc flags. `pi-codegraph` pins the binary, gates which repos it runs against, and gives the agent one stable, scriptable surface.

Install

Requires **Node ≥ 18**. The wrapped engine is separate:

git clone https://github.com/renezander030/pi-codegraph
ln -s "$PWD/pi-codegraph/pi-codegraph" /usr/local/bin/pi-codegraph

# install the engine (see its repo), then point pi-codegraph at it:
pi-codegraph config-bin /path/to/codebase-memory-mcp   # or set PI_CODEGRAPH_BIN
pi-codegraph doctor                                     # confirm it's found

Quick start

pi-codegraph doctor                       # is the engine installed?
pi-codegraph trust --repo . --label