codemap — a project brain for your AI. Give LLMs instant architectural context without burning tokens.
# macOS/Linux
brew tap JordanCoin/tap && brew install codemap
# Windows
scoop bucket add codemap https://github.com/JordanCoin/scoop-codemap
scoop install codemapOther options: Releases |
go install| Build from source
If you install codemap from a release tarball, also install ast-grep separately for --deps.
The tarball includes codemap and the bundled rules, but not the ast-grep executable.
Example for Alpine-based CI:
apk add --no-cache curl jq bash python3 py3-pip
ARCH=$(uname -m)
if [ "$ARCH" = "x86_64" ]; then ARCH="amd64"; elif [ "$ARCH" = "aarch64" ]; then ARCH="arm64"; fi
CODEMAP_VERSION=$(curl -fsSL https://api.github.com/repos/JordanCoin/codemap/releases/latest | jq -r '.tag_name' | tr -d 'v')
curl -fsSL "https://github.com/JordanCoin/codemap/releases/download/v${CODEMAP_VERSION}/codemap_${CODEMAP_VERSION}_linux_${ARCH}.tar.gz" \
| tar xz -C /usr/local/bin/ codemap
python3 -m pip install --no-cache-dir ast-grep-cliIf you want a self-contained archive for CI/CD, use the codemap-full release artifact instead.
It includes codemap, ast-grep, and sg in one archive so --deps works after extraction.
apk add --no-cache curl jq bash
ARCH=$(uname -m)
if [ "$ARCH" = "x86_64" ]; then ARCH="amd64"; elif [ "$ARCH" = "aarch64" ]; then ARCH="arm64"; fi
CODEMAP_VERSION=$(curl -fsSL https://api.github.com/repos/JordanCoin/codemap/releases/latest | jq -r '.tag_name' | tr -d 'v')
curl -fsSL "https://github.com/JordanCoin/codemap/releases/download/v${CODEMAP_VERSION}/codemap-full_${CODEMAP_VERSION}_linux_${ARCH}.tar.gz" \
| tar xz -C /usr/local/bin/ codemap ast-grep sgNo repo clone is required for normal users. Run setup from your git repo root (not a subdirectory), or hooks may not resolve project context.
# install codemap first (package manager)
brew tap JordanCoin/tap && brew install codemap
# then run setup inside your project
cd /path/to/your/project
codemap setupcodemap setup is the default onboarding path and configures the pieces that make codemap most useful with Claude:
- creates
.codemap/config.json(if missing) with auto-detected language filters - installs codemap hooks into
.claude/settings.local.json(project-local by default) - hooks automatically start/read daemon state on session start
Use global Claude settings instead of project-local settings:
codemap setup --globalWindows equivalent:
scoop bucket add codemap https://github.com/JordanCoin/scoop-codemap
scoop install codemap
cd C:\path\to\your\project
codemap setupOptional helper scripts (mainly for contributors running from this repo):
- macOS/Linux:
./scripts/onboard.sh /path/to/your/project - Windows (PowerShell):
./scripts/onboard.ps1 -ProjectRoot C:\path\to\your\project
- Restart Claude Code or open a new session.
- At session start, you should see codemap project context.
- Edit a file and confirm pre/post edit hook context appears.
codemap . # Fast tree/context view (respects .codemap/config.json)
codemap --diff # What changed vs main
codemap handoff . # Save layered handoff for cross-agent continuation
codemap --deps . # Dependency flow (requires ast-grep)
codemap skill list # Show available skills
codemap context # Universal JSON context for any AI tool
codemap mcp # Run Codemap MCP server on stdio
codemap plugin install # Install the Codemap Codex plugin
codemap serve # HTTP API for non-MCP integrationscodemap --only swift .
codemap --exclude .xcassets,Fonts,.png .
codemap --depth 2 .
codemap github.com/user/repo| Flag | Description |
|---|---|
--depth, -d <n> |
Limit tree depth (0 = unlimited) |
--only <exts> |
Only show files with these extensions |
--exclude <patterns> |
Exclude files matching patterns |
--diff |
Show files changed vs main branch |
--ref <branch> |
Branch to compare against (with --diff) |
--deps |
Dependency flow mode |
--importers <file> |
Check who imports a file |
--skyline |
City skyline visualization |
--animate |
Animate the skyline (use with --skyline) |
--json |
Output JSON |
Note: Flags must come before the path/URL:
codemap --json github.com/user/repo
Smart pattern matching — no quotes needed:
.png→ any.pngfileFonts→ any/Fonts/directory*Test*→ glob pattern
See what you're working on:
codemap --diff
codemap --diff --ref develop╭─────────────────────────── myproject ──────────────────────────╮
│ Changed: 4 files | +156 -23 lines vs main │
╰────────────────────────────────────────────────────────────────╯
├── api/
│ └── (new) auth.go ✎ handlers.go (+45 -12)
└── ✎ main.go (+29 -3)
⚠ handlers.go is used by 3 other files
See how your code connects:
codemap --deps .╭──────────────────────────────────────────────────────────────╮
│ MyApp - Dependency Flow │
├──────────────────────────────────────────────────────────────┤
│ Go: chi, zap, testify │
╰──────────────────────────────────────────────────────────────╯
Backend ════════════════════════════════════════════════════
server ───▶ validate ───▶ rules, config
api ───▶ handlers, middleware
HUBS: config (12←), api (8←), utils (5←)
codemap --skyline --animateAnalyze any public GitHub or GitLab repo without cloning it yourself:
codemap github.com/anthropics/anthropic-cookbook
codemap https://github.com/user/repo
codemap gitlab.com/user/repoUses a shallow clone to a temp directory (fast, no history, auto-cleanup). If you already have the repo cloned locally, codemap will use your local copy instead.
18 languages for dependency analysis: Go, Python, JavaScript, TypeScript, Rust, Ruby, C, C++, Java, Swift, Kotlin, C#, PHP, Bash, Lua, Scala, Elixir, Solidity
Powered by ast-grep. Install via
brew install ast-grepfor--depsmode.
If you want a compact review bundle for another LLM, combine the three high-signal views:
codemap --json --diff --ref main .
codemap --json --deps --diff --ref main .
codemap --json --importers path/to/file .For a reusable built-in command that emits either Markdown, text, or a single JSON object:
codemap blast-radius --ref main .
codemap blast-radius --json --ref main .
codemap blast-radius --text --ref main .Hooks (Recommended) — Automatic context at session start, before/after edits, and more. → See docs/HOOKS.md
MCP Server — Deep integration with project analysis + handoff tools. → See docs/MCP.md
codemap now supports a shared handoff artifact so you can switch between agents (Claude, Codex, MCP clients) without re-briefing.
codemap handoff . # Build + save layered handoff artifacts
codemap handoff --latest . # Read latest saved artifact
codemap handoff --json . # Machine-readable handoff payload
codemap handoff --since 2h . # Limit timeline lookback window
codemap handoff --prefix . # Stable prefix layer only
codemap handoff --delta . # Recent delta layer only
codemap handoff --detail a.go . # Lazy-load full detail for one changed file
codemap handoff --no-save . # Build/read without writing artifactsWhat it captures (layered for cache reuse):
prefix(stable): hub summaries + repo file-count contextdelta(dynamic): changed file stubs (path,hash,status,size), risk files, recent events, next steps- deterministic hashes:
prefix_hash,delta_hash,combined_hash - cache metrics: reuse ratio + unchanged bytes vs previous handoff
Artifacts written:
.codemap/handoff.latest.json(full artifact).codemap/handoff.prefix.json(stable prefix snapshot).codemap/handoff.delta.json(dynamic delta snapshot).codemap/handoff.metrics.log(append-only metrics stream, one JSON line per save)
Save defaults:
- CLI saves by default; use
--no-saveto make generation read-only. - MCP does not save by default; set
save=trueto persist artifacts.
Compatibility note:
- legacy top-level fields (
changed_files,risk_files, etc.) are still included for compatibility and will be removed in a future schema version after migration.
Why this matters:
- default transport is compact stubs (low context cost)
- full per-file context is lazy-loaded only when needed (
--detail/file=...) - output is deterministic and budgeted to reduce context churn across agent turns
Hook integration:
session-stopwrites.codemap/handoff.latest.jsonsession-startshows a compact recent handoff summary (24h freshness window)
CLAUDE.md — Add to your project root to teach Claude when to run codemap:
cp /path/to/codemap/CLAUDE.md your-project/Set per-project defaults in .codemap/config.json so you don't need to pass --only/--exclude/--depth every time. Hooks also respect this config.
codemap config init # Auto-detect top extensions, write config
codemap config show # Display current configExample .codemap/config.json:
{
"only": ["rs", "sh", "sql", "toml", "yml"],
"exclude": ["docs/reference", "docs/research"],
"guidance": {
"missing_extension_hints": true,
"ignored_extensions": []
},
"depth": 4,
"mode": "auto",
"budgets": {
"session_start_bytes": 30000,
"diff_bytes": 15000,
"max_hubs": 8
},
"routing": {
"retrieval": { "strategy": "keyword", "top_k": 3 },
"subsystems": [
{
"id": "watching",
"paths": ["watch/**"],
"keywords": ["hook", "daemon", "events"],
"docs": ["docs/HOOKS.md"],
"agents": ["codemap-hook-triage"]
}
]
},
"drift": {
"enabled": true,
"recent_commits": 10,
"require_docs_for": ["watching"]
}
}When an MCP file search has no visible results but finds real matches hidden by only, Codemap reports the matching paths and suggests which extensions to include. These hints still respect exclude and never modify project config. Set guidance.missing_extension_hints to false to disable all hints, or list extensions in guidance.ignored_extensions to suppress only those suggestions.
All fields are optional. CLI flags always override config values. Hook-specific policy fields are optional and bounded by safe defaults.
codemap ships with a skills framework — markdown files that provide context-aware guidance to AI agents. Skills are automatically matched against your intent, the files you mention, and the languages in your project.
codemap skill list # Show all available skills
codemap skill show hub-safety # Print full skill content
codemap skill init # Create a custom skill template| Skill | Activates When |
|---|---|
hub-safety |
Editing hub files (3+ importers) |
refactor |
Restructuring, renaming, moving code |
test-first |
Writing tests, TDD workflows |
explore |
Understanding how code works |
handoff |
Switching between AI agents |
Drop a .md file in .codemap/skills/ with YAML frontmatter:
---
name: my-skill
description: When this skill should activate
keywords: ["relevant", "keywords"]
languages: ["go"]
---
# Instructions for the AI agentProject-local skills override builtins. No Go code needed — just markdown.
Skills are also available via MCP: list_skills (metadata) and get_skill (full body).
The prompt-submit hook performs intent classification on every prompt — detecting whether you're refactoring, fixing a bug, exploring, testing, or building a feature. It then:
- Surfaces risk analysis based on hub file involvement
- Shows your working set (files edited this session)
- Emits structured JSON markers (
<!-- codemap:intent -->) for tool consumption - Matches relevant skills and tells you which to pull (
codemap skill show <name>) - Warns about documentation drift when docs are stale
A single command that gives any AI tool codemap's full intelligence:
codemap context # Full JSON envelope
codemap context --for "refactor auth" # With pre-classified intent + matched skills
codemap context --compact # Minimal for token-constrained agentsThe output is a ContextEnvelope containing project metadata, intent classification, working set, matched skills, and handoff reference. Cursor, Windsurf, Codex, custom agents — anything that can shell out gets code-aware intelligence.
For tools that prefer HTTP over CLI:
codemap serve --port 9471| Endpoint | Returns |
|---|---|
GET /api/context?intent=refactor+auth |
Full context envelope |
GET /api/context?compact=true |
Minimal envelope |
GET /api/skills |
All skills with metadata |
GET /api/skills?language=go&category=refactor |
Filtered skill matches |
GET /api/skills/<name> |
Full skill body |
GET /api/working-set |
Current session's active files |
GET /api/health |
Server health check |
Binds to 127.0.0.1 by default. Use --host 0.0.0.0 to expose to network.
When you switch between AI agents (Claude → Codex → Cursor), codemap tracks who worked and what they did:
{
"agent_history": [
{"agent_id": "claude-code", "files_edited": ["cmd/hooks.go", "main.go"], "ended_at": "..."},
{"agent_id": "codex", "files_edited": ["scanner/types.go"], "ended_at": "..."}
]
}Agent detection is automatic via environment variables. History is carried across sessions (capped at 20 entries) in the handoff artifact.
- Diff mode, Skyline mode, Dependency flow
- Tree depth limiting (
--depth) - File filtering (
--only,--exclude) - Project config (
.codemap/config.json) - Claude Code hooks & MCP server
- Cross-agent handoff artifact (
.codemap/handoff.latest.json) - Remote repo support (GitHub, GitLab)
- Intelligent routing (intent classification, risk analysis, working set)
- Skills framework (builtin + custom skills, CLI, MCP tools)
- Context protocol (
codemap context— universal JSON envelope for any AI tool) - HTTP API (
codemap serve— REST endpoints for non-MCP integrations) - Agent-aware handoff (multi-agent history tracking)
- Community skill registry (GitHub-hosted,
codemap skill add <name>) - Enhanced analysis (entry points, key types)
- Fork → 2. Branch → 3. Commit → 4. PR
MIT

