AgentStack
MCP verified MIT Self-run

Pi Codegraph

mcp-renezander030-pi-codegraph · by renezander030

Pi-agent wrapper around codebase-memory-mcp: derived code knowledge (call graphs, blast radius) over a harness-owned trust boundary

No reviews yet
0 installs
2 views
0.0% view→install

Install

$ agentstack add mcp-renezander030-pi-codegraph

✓ scanned · ✓ verified — works with Claude Code, Cursor, and more.

Security review

✓ Passed

No issues found. Passed automated security review. · v0.1.0 How review works →

  • Prompt-injection patterns
  • Secret / credential exfiltration
  • Dangerous shell & filesystem operations
  • Untrusted network calls
  • Known-malicious package signatures

What it can access

  • Network access No
  • Filesystem access No
  • Shell / process execution No
  • Environment & secrets No
  • Dynamic code execution No

From automated source analysis of v0.1.0. “Used” means the capability is present in the source — more access means more to trust, not that it’s unsafe.

Are you the author of Pi Codegraph? Claim this listing to set pricing, connect Stripe payouts, and keep 70% of every sale.
Sign up to claim

About

pi-codegraph

A trusted Pi-side wrapper around 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 (which carries authored knowledge) and a sibling of 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 app   # register this repo as a codegraph source
pi-codegraph index --repo .               # parse the repo into the graph (run once, before querying)
pi-codegraph arch --repo . -H             # architecture overview
pi-codegraph trace ProcessOrder --direction inbound   # who calls it
pi-codegraph impact                       # blast radius of the current git diff
pi-codegraph search "auth middleware"     # graph search (search_graph, BM25 + line ranges)
pi-codegraph query trace_path --args '{"function_name":"main","depth":3}'

The engine keys each project by a slug of its absolute path; the convenience commands (and query) resolve that project from --repo for you, so you never hand-pass it. index must run before any query — the engine has no data otherwise.

plan shows the exact JSON-RPC it would send without launching anything:

pi-codegraph plan get_architecture --repo . -H

Pi Extension

This repo does not ship the Pi extension surface directly. Use the standalone pi-recall package when you want Pi slash commands such as /recall, /recall-arch, and /recall-impact.

pi-recall delegates to this CLI and keeps this repo's trust gate intact.

CLI

| Command | Purpose | |---|---| | doctor | Resolve the binary (env / config / PATH); report presence + version. | | config-bin | Pin the engine binary path. | | trust [--label L] | Register a repo as a trusted codegraph source. | | repos | List trusted repos. | | index | Parse the repo into the knowledge graph (run before any query). | | tools | List the live server's tools. | | trace [--direction] [--depth] | Call chain into/out of a function. | | arch / impact / schema | Architecture, git-diff blast radius, graph schema. | | search [--limit N] / snippet | Graph search (search_graph); source for one symbol. | | query [--args JSON] | Call any of the engine's tools directly. | | plan [--args JSON] | Print the JSON-RPC request (no spawn). |

JSON by default; -H/--human for a table.

Trust model

Faithful to pi-gate:

| Invariant | How | |---|---| | The third-party binary is untrusted | we only resolve a path to it (env PI_CODEGRAPH_BIN > ~/.config/pi-codegraph/config.json > PATH) and speak MCP over stdio | | The record of which repos are queryable lives outside the agent tree | trusted-repo registry at ~/.config/pi-codegraph/repos.json, state at ~/.local/state/pi-codegraph/ | | Queries run only on a registered repo | the gate blocks unregistered repos; --override "" bypasses with an audit trail | | Stable repo identity across moves | git root-commit sha, falling back to a path hash |

How it talks to the engine

MCP stdio transport: spawn the server with cwd = the repo, initialize, notifications/initialized, then tools/list or tools/call, reading newline-delimited JSON-RPC. The test suite proves the full handshake against a mock server, so the wrap is verified without the Go binary installed.

Roadmap

Auto-index on first trust, a streaming mode for large results, per-tool allow/deny in the registry, and a paired pi-okf exporter that turns the derived graph into an OKF bundle (derived knowledge seeding authored).

License

MIT

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

Install and usage instructions live in the source repository linked above.

Reviews

No reviews yet — be the first.

Versions

  • v0.1.0 Imported from the upstream source.