Phase 2 · Connectivity·12 min·6 steps

Code intelligence — codebase-memory-mcp + codegraph

Index your repo into a knowledge graph so Claude can answer 'who calls this function' and 'what breaks if I delete it' in one tool call instead of grepping.

6 steps0%

Du liest ohne Account. Mit Login speichern wir Step-Fortschritt + Notes.

Code intelligence

Two MCP servers turn Claude from "smart grep" into "engineer who knows your codebase":

codebase-memory-mcp — indexes your repo into a Postgres-backed knowledge graph. Symbols, files, edges (calls, imports, defines).
codegraph — runs deep analysis on top: complexity, find_cycles, fn_impact, audit, communities.

Together they replace 80% of grep + ad-hoc code reading with O(1) graph queries.

The graph runs locally — your code never leaves your machine. The Postgres index sits in .codebase-memory/graph.db.zst next to your repo.

Step 1: Install codebase-memory-mcp

claude mcp add codebase-memory -s user -- npx -y codebase-memory-mcp

For Codex:

[mcp_servers.codebase-memory]
command = "npx"
args = ["-y", "codebase-memory-mcp"]

Step 2: Install codegraph (optional but recommended)

claude mcp add codegraph -s user -- codegraph mcp

If codegraph is not on PATH, install with:

npm install -g codegraph

Codegraph is the "deep analysis" layer. You can skip it and just use codebase-memory-mcp's basic search; you lose fn_impact and find_cycles etc.

Step 3: Index your first repo

In a fresh Claude Code session inside your repo:

"index this repository"

Claude calls index_repository on codebase-memory-mcp. First index is ~1-2 minutes for a 50k-line repo. Subsequent indexes are incremental (content-hashed) — sub-second.

Then ask:

"what's the architecture overview of this repo?"

Claude calls get_architecture — you get module hierarchy, dependency graph, and entry points. This single call replaces 30 minutes of click-around.

Step 4: Verify

Run aiguide_validate_step. The validator looks for any of codebase|codegraph|archtracker in claude mcp list, OR a .codebase-memory/ directory in the cwd. Either path passes.

Step 5: Useful queries to try

Once indexed, your AI can answer:

"Who calls handleStartRecipe?" → search_graph returns all callers in O(1).
"What breaks if I rename this function?" → codegraph fn_impact returns transitive callers.
"Are there circular dependencies?" → codegraph find_cycles returns SCCs.
"What are the most complex files?" → codegraph complexity returns hotspots.
"Show me the call chain from POST /upgrade to Stripe" → trace_call_path.

Add a CLAUDE.md hint so Claude reaches for the graph instead of grep:

## Codebase intelligence

When working in this repo, prefer search_graph / codegraph context / fn_impact
over grep + reading whole files. Re-index after structural changes
(new files, renamed functions). The graph is faster and more accurate
than text search for "who calls X" and "what depends on Y" questions.

Client check · run on your machine

(claude mcp list 2>&1 | grep -iE "codebase|codegraph|archtracker") || ls -d ./.codebase-memory 2>/dev/null || echo "neither code-intel MCP nor local index"

Expect: Either a code-intel MCP is listed OR a `.codebase-memory/` directory exists in your project.

If stuck: Install codebase-memory-mcp (`claude mcp add codebase-memory -s user -- npx -y codebase-memory-mcp`) and run `index_repository` once.

Step 6: Optional — archtracker for drift detection

claude mcp add archtracker -s user -- archtracker-mcp

archtracker takes architecture snapshots and tells you when the codebase drifts (orphan modules, coupling hotspots, circular deps that didn't exist last week). Mostly useful for repos with multiple contributors — solo projects don't drift fast enough to need it.

After this recipe Claude understands your code structure as well as you do. The next recipe wires up GitHub so it can read PRs and issues too.

← Web research — Brave, Exa, and GitHub MCP — read PRs, issues, →