Code Graph
To find who calls a function, an agent usually greps, opens a file, greps again — and reads thousands of tokens of code it didn’t need. Draw an arrow from an agent to the Code Graph card and it asks a graph of your project instead: callers, callees, definitions, in one call.
Built on codebase-memory-mcp, the open-source code graph engine by DeusData:github.com/DeusData/codebase-memory-mcp
12
graph tools an agent gets with one arrow — search, callers and callees, source, Cypher, architecture
≈5
seconds to update the graph after an edit, in our test — only what changed is re-indexed
0
bytes of your code sent anywhere — the engine runs on your computer, with no telemetry
One question, answered by the graph
The calls and their results are the ones the real app passed to the model in a test run, on a copy of this website’s own code.
❯ Who calls localHref? I want to change its signature.
843
symbols
73 files · 423 calls
Up to date· TypeScript
Last call
Shortened for the page: the search returned 31 matching functions. The agent’s answer is illustrative.
The engine’s own tools, scoped to your workspace
The same tools and descriptions as codebase-memory-mcp’s MCP server, prefixed like every NeuroSquad plugin tool. NeuroSquad fills in the project, so an agent only ever reaches its own workspace’s graph.
Find
codegraph_search_graph finds functions, classes and other symbols by name, pattern or meaning; codegraph_search_code is text search ranked by the graph.
Follow
codegraph_trace_path shows who calls a function and what it calls, several hops deep; codegraph_query_graph answers multi-step questions in read-only Cypher.
Read
codegraph_get_code_snippet returns the source of one symbol, codegraph_get_file_outline everything declared in a file — without opening whole files.
Understand
codegraph_get_architecture lists languages, packages, entry points and hotspots; codegraph_detect_changes maps your uncommitted changes to the symbols they affect.
The graph follows your edits
A stale map is worse than none. The card keeps the graph in step with the files on its own.
Watches the folder
When a file changes, the card marks the graph out of date and re-indexes a few seconds later. Dependency and build folders are ignored.
Incremental
Only what changed is parsed again, so an update takes seconds even when the first build of a large project took minutes.
After every turn
When a connected agent finishes a turn, the graph is refreshed if anything changed — the next question sees the new code.
Honest answers
While an update runs, the agent’s answer says so, and the card shows Updating. Agents can also re-index on demand.
One card, one arrow
Everything happens on the canvas you already have open.
Add the card
Add-card menu → “Plugin…” → Code Graph → “Add to canvas”.
Download and build
Press Download once (about 40 MB), then Build graph. Most projects take seconds.
Draw an arrow
From an agent to the card. It gets the graph tools on its next tool listing — no restart for most CLIs.
Ask structural questions
“Who calls this?”, “what breaks if I change that?” — the agent asks the graph. The card shows which agent asked what.
Every CLI NeuroSquad runs
The tools come through NeuroSquad’s own MCP server, so every agent with an arrow can use them.
The graph tools — every agent NeuroSquad connects to its MCP server
Workspaces on this computer, in WSL and on SSH hosts (Linux x86-64 / ARM64): the engine runs where the code is.
Your code stays on your computer
The engine is a program on your machine, not a service.
Nothing sent
codebase-memory-mcp makes no network requests and collects no telemetry. Indexing and every query run locally.
Verified download
The engine is downloaded once from its official GitHub release and checked against a checksum pinned in the app. A file that doesn’t match is deleted.
Kept in the app’s folder
The engine and the graph live in NeuroSquad’s data folder (for a WSL or SSH workspace, the app’s own folder there) — nothing is written into your project or your home folder.
Your own install untouched
If you already use codebase-memory-mcp yourself, it keeps working separately: NeuroSquad runs its copy apart from it.
Open source inside
Powered by codebase-memory-mcp
The indexing, the graph and the tools are codebase-memory-mcp’s: tree-sitter parsing for over a hundred languages, type-aware call resolution, a SQLite graph and Cypher queries, by DeusData under the MIT license. NeuroSquad downloads its official build and runs it for your workspaces.
Questions about Code Graph
Agents that know your code’s structure
Download NeuroSquad, add the Code Graph card, and draw an arrow.