Local-first code intelligence

Your agent greps.
codemap knows.

Give coding agents a cited map of definitions, callers, blast radius, and tests—before they change the code. One local binary, exposed over MCP and CLI.

Set up your agent Read the quick start

brew install abdul-hamid-achik/tap/codemap

Pure-Go binary
Stored-graph queries work offline
7 indexed languages
42 MCP tools in full

structural evidence cited

codemap_impact("Store.NodeAtLine")

callers7 direct

callees4 edges

verification13 tests

blast radius23 symbols

exact definitionStore.NodeAtLineinternal/graph/queries.go

call graph resolved

freshness current

selectordurable

1 callevery result carries file:line

Measured on a real repository

Less searching. More verified work.

A hermetic 60-session A/B on go-git, modelled with the shipped core profile and agent playbook. Directional results, negative runs included.

Read the benchmark methodology →

44%: fewer tool calls
40%: faster completion
13%: lower modelled cost
60: isolated sessions

One graph, three surfaces

Ask once. Follow exact evidence.

codemap joins semantic intent to structural relationships, then preserves the evidence an agent needs to act safely.

01
Find the intent
Search by name, exact text, or meaning. Every hit lands on a real symbol.
semantic "jwt validation"
02
Expand the structure
Join definitions, calls, callback references, imports, and annotations.
context authenticateUser
03
Verify the change
Return blast radius, covering tests, and confidence before code is changed.
review --fail-on-risk high

Built for agent loops

A code map that earns trust.

Fast enough for every edit, explicit enough for automation, and useful even when embeddings or language servers are unavailable.

01 · contextone call

Everything around a symbol, already joined.

Definition, source, callers, callees, value references, blast radius, and covering tests arrive as one bounded response.

› codemap_context("Authorize")

definition: auth/service.go:84
callers: 6 exact edges
covering tests: 8 selected

02 · confidencemachine-readable

Honest when the graph is incomplete.

No vague “best effort.” Call-graph and impact reports carry explicit confidence signals.

resolvedexact coverage
namecandidate edges
unresolvedreindex action

03 · reviewdiff aware

Turn every edit into a test plan.

Map the current diff to affected code, untested hotspots, and one risk band.

aggregate riskmedium · 0.54

12 affected symbols · 5 covering tests

04 · retrievalhybrid

Search by meaning. Land on structure.

Vector + BM25 retrieval hands durable selectors to the graph—not loose snippets.

0.92

auth.TokenValidator.Validateinternal/auth/token.go:37

0.87

middleware.requireJWTapi/middleware/auth.go:52

05 · local firstdegrades cleanly

Your code stays local by default.

Structural queries work offline after indexing. Semantic retrieval can use local Ollama—or an explicitly configured remote endpoint.

graphSQLite · pure Go
vectorsveclite · local
transportstdio MCP

The difference, in one question

“What breaks if I change `Store.NodeAtLine`?”

agent without codemapguessing

● Grep("NodeAtLine")
● Read(internal/graph/queries.go)
● Grep("NodeAtLine(")
● Read(internal/app/service_grep.go)
● Read(internal/app/symbol_at_batch.go)
… 19 more greps and reads

agent with codemapcited

● codemap_impact("Store.NodeAtLine")
{
  "direct_callers": 7,
  "blast_radius": 23,
  "covering_tests": 13,
  "call_graph": "resolved"
}

Illustrative real-repository response shape. The measured benchmark above contains the reproducible aggregate result and methodology.

Fits the loop you already use

From local agent to merge gate.

Install the same structural workflow in your coding harness and in pull-request automation.

One command per harness.

Register the MCP server without clobbering existing configuration, then install a playbook that teaches the agent when to use codemap.

$codemap agent setup claude-code

claude-code
cursor
codex
gemini
cline
roo
zed
vscode
opencode
aider
agents-md

See agent integrations →

Review every pull request.

Post one sticky impact report and fail only on the risk or test-coverage rules your team chooses.

jobs:
  review:
    uses: abdul-hamid-achik/codemap/
      .github/workflows/codemap-review-reusable.yml@main
    with:
      fail-on-untested: 'true'
      fail-on-risk: 'high'

Configure the CI review gate →

Start with one repository

Give your agent the map.

Install codemap, index the repository, and ask your first structural question.

Index your first project Explore all MCP tools

1codemap init

2codemap index --precise

3codemap context YourSymbol

Your agent greps.codemap knows.