agentmap

The repo map your coding agent is forced to use — ~98% fewer context tokens to understand your TS/JS codebase.

Your AI coding agent re-learns your codebase every session — opening files and grepping to find what connects to what, burning tokens before it writes a line. agentmap gives it a queryable, ranked code-relationship map for TypeScript/JavaScript repos instead — a ts-morph import/symbol graph ranked by personalized PageRank. Ask it to "add a field" or "fix the login bug" and it finds the right files, their imports, and what already exists in ~98% fewer context tokens on average (up to ~99.9% per task; figures are chars/4 estimates applied equally to both sides) — kept current by a post-commit auto-refresh and actually used via a PreToolUse(Grep) hook.

One file, one runtime dependency (ts-morph, which bundles the TypeScript compiler — ~10 MB installed). No vector DB, no embedding API, no server. npx @raymondchins/agentmap --any <query> and you have a ranked answer.

Benchmark

Every task you hand a coding agent starts with the same hidden step — find the relevant code. Here's the token cost of that step, reading raw files vs querying agentmap, on a real 154-file Next.js app (vercel/ai-chatbot). Every figure is captured tool output (node benchmark/bench.mjs <repo> at the pinned sha):

Context tokens the agent burns to answer each question — token est = chars/4, applied to both sides.

That's the agent reaching the same answer on 58× fewer tokens overall — and the pattern holds across zod (367 files, 99.2%) and taxonomy (125 files, 96.0%), peaking at 646× fewer on a single whole-repo map. Reproducible at pinned shas; full per-scenario tables in ./benchmark/RESULTS.md.

Methodology note: the 58× overall figure is dominated by the whole-repo-load scenario (Scenario F — 150 K vs 1 K tokens), which skews the combined ratio sharply upward. Excluding it, the per-task overall ratio on the same sample repo is approximately 32×. Both numbers are real; the headline captures the most common agent worst-case (repo-dump on session start), while the per-task average better represents typical individual queries. RESULTS.md has the full breakdown.

Fewer tokens, but are they the right tokens? Token efficiency is only half the story — a separate EVAL.md (npm run eval) scores retrieval accuracy against ground truth derived live from real repos (zod, zustand, hono). Headline: agentmap returns the symbol definition in the top 3 ~95% of the time (naive grep ~79%) at ~2.6× fewer tokens, and identifies a module's dependents at ~100% precision (grep ~58%). Honest tradeoffs and method in EVAL.md.

Speed: a cold build (parse + PageRank + symbol graph) takes ~1.2s; a warm cached query returns in ~0.1s (the lazy-loaded path added in 0.2.2) — the agent has a ranked answer back before it would have finished opening the first handful of files.

Honest notes: the win scales with the work — the small rows above (63%, 11%) are the floor, and a trivial single-file lookup can even cost more than cat+grep (taxonomy's file-import task hit −313%; we leave it in). Numbers measure context-token volume, not answer quality or wall-clock.

Why it's different

Most "repo context" tools are a photocopy: they dump your repository (or a slice of it) into the prompt once and walk away. The copy goes stale the moment you edit a file, and nothing makes the agent actually read it.

agentmap is the opposite — a queryable, ranked, self-refreshing map the agent interrogates flag-by-flag, that rebuilds itself on every commit, and that a PreToolUse hook steers the agent toward before it falls back to serial grep.

| | agentmap | Aider repo map | RepoMapper | Repomix | code2prompt | | --- | --- | --- | --- | --- | --- | | Ranking algorithm | Personalized PageRank (file + symbol graphs) | PageRank (graph ranking) | Importance heuristics | None (file order) | None (file order) | | Languages | TS/JS (via ts-morph) | Many (tree-sitter) | Many (tree-sitter) | Language-agnostic (text) | Language-agnostic (text) | | Token-budget output | Yes — --map [--tokens N] ranked digest | Yes (built into Aider's context) | Partial | Yes (size caps) | Yes (templates/caps) | | Agent-loop integration | Yes — post-commit auto-refresh + PreToolUse hook | In-process (Aider only) | No | No | No | | Dependencies | ts-morph only | Python + tree-sitter stack | Python + tree-sitter | Node | Rust binary | | Install | npx @raymondchins/agentmap | pip install aider | pip install | npx/global | cargo/binary |

What that table is not claiming: agentmap is TS/JS-only (the others are multi-language), and it's a file-level import graph, not a full call-site/reference resolver (see Scope & limitations). The differentiators are narrow and honest: (1) the --any router, and (2) the agent-loop wiring. Everything else is table stakes.

The agent loop (the actual point)

Here's the quiet failure of every other repo-map tool: it builds a beautiful map, and then the agent forgets it exists and greps anyway. A map the agent doesn't open is just dead weight.

agentmap closes that loop. Two hooks (in ./hooks/) do the work: the map refreshes itself after every commit, and the agent gets nudged to query it before it serial-greps. You wire it once — then it stays current on its own, and stays used.

1. Auto-refresh on commit

hooks/post-commit rebuilds .claude/agentmap.json after each commit, detached + silenced so it never slows the commit. It skips during rebase/merge/cherry-pick and no-ops if Node is missing.

The hooks ship inside the npm package. The simplest setup:

npx @raymondchins/agentmap --install-hooks

This copies hooks/post-commit into .git/hooks/, sets it executable, ensures .claude/agentmap.json is in .gitignore, and auto-wires the PreToolUse nudge hook into .claude/settings.json (merge-safe + idempotent) so map enforcement is on by default — no manual paste. Manual alternative for just the post-commit hook:

# from your repo root
cp hooks/post-commit .git/hooks/post-commit
chmod +x .git/hooks/post-commit

The hook auto-locates the builder: a local agentmap.mjs, then scripts/agentmap.mjs, then the installed agentmap binary, then npx --no-install @raymondchins/agentmap.

2. Force the agent to use it — PreToolUse hook

hooks/agentmap-nudge.mjs is a non-blocking hook for Claude Code that covers both the Grep tool and raw Bash text-searchers (grep/rg/egrep/fgrep/ag/ack). When either looks like a dependency / who-imports / component-usage / reuse / where-is-symbol search, it injects a reminder steering the agent to agentmap --any first. It never denies the call, and stays silent for raw-string / Tailwind-class / lowercase-HTML-tag sweeps and for pipe-filtered commands like ps aux | grep node — so it's high-signal, not nagging.

Fires on: import/require/export/from '...' patterns, JSX component tags (<Hero, <ProviderCard), explicit intent words (where is, who imports, reuse, existing component), and — in the Bash branch — bare multi-hump PascalCase identifiers (ProviderCard, TopProviders) that almost always mean "where is this symbol / who uses it". The Bash branch only fires when the searcher is the primary command (at the start, or after ;/&&); piped log-filters stay silent.

--install-hooks writes both matchers into .claude/settings.json for you (merge-safe — preserves existing settings, won't duplicate on re-run). The single hook file dispatches internally on tool_name. For reference, or to wire it by hand:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Grep",
        "hooks": [{ "type": "command", "command": "node ./hooks/agentmap-nudge.mjs" }]
      },
      {
        "matcher": "Bash",
        "hooks": [{ "type": "command", "command": "node ./hooks/agentmap-nudge.mjs" }]
      }
    ]
  }
}

That's the "forced to use it" in the tagline: the map stays current on its own, and the agent is steered to it the moment it reaches for a dependency-shaped grep or Bash search.

3. Agent skills (Cursor, Claude Code, Codex, OpenCode, Gemini, Antigravity, Copilot)

npx @raymondchins/agentmap --install-skill

Copies packaged SKILL.md files and a Cursor rule (.cursor/rules/agentmap.mdc, alwaysApply: true) into the current repo or global agent directories. Paths follow each platform's official skill-directory conventions. Options:

agentmap --install-skill --platform cursor           # Cursor rule only (project)
agentmap --install-skill --platform claude           # .claude/skills/agentmap/SKILL.md
agentmap --install-skill --platform codex            # .codex/skills/ (project) or ~/.codex/skills/ (global)
agentmap --install-skill --platform opencode         # .opencode/skills/ (project) or ~/.config/opencode/skills/ (global)
agentmap --install-skill --platform gemini           # .gemini/skills/ (project); global ~/.gemini/skills/ (Windows global: ~/.agents/skills/)
agentmap --install-skill --platform antigravity      # .agents/skills/ (project) or ~/.gemini/config/skills/ (global)
agentmap --install-skill --platform copilot          # .copilot/skills/ or ~/.copilot/skills/
agentmap --install-skill --global --platform claude  # ~/.claude/skills/...
agentmap --install-skill --platform agents           # legacy .agents/skills/ (project or global); excluded from default `all`
agentmap --install-skill --dry-run                   # preview paths, no writes

--platform all installs: claude, cursor, codex, opencode, gemini, antigravity, copilot (not legacy agents).

Some platforms also get always-on docs and hooks in the same command:

| --platform | Skill | Also installs (project) | Global docs | |--------------|-------|-------------------------|-------------| | gemini | .gemini/skills/…/SKILL.md | GEMINI.md + .gemini/settings.json BeforeTool nudge | ~/.gemini/GEMINI.md | | codex | .codex/skills/…/SKILL.md | AGENTS.md merge-safe <!-- agentmap:begin/end --> block | ~/.codex/AGENTS.md | | opencode | .opencode/skills/…/SKILL.md | AGENTS.md + .opencode/plugins/agentmap-nudge.js | ~/.config/opencode/AGENTS.md |

Codex and OpenCode share one repo-root AGENTS.md on project install. Existing content outside the marked block is preserved.

Pair with --install-hooks (Claude Code) or --mcp (Cursor MCP).

Quickstart

No install needed:

npx @raymondchins/agentmap --any <query>

…or run it directly from a checkout:

node agentmap.mjs --any <query>

The first run builds and caches the map to .claude/agentmap.json (add it to .gitignore). Subsequent runs serve the cache when the tree is clean and HEAD is unchanged, and silently rebuild from disk when there are uncommitted .ts/.tsx/.js/... edits — so queries always reflect your in-flight work.

Run with no flag to build + print a one-line summary:

$ node agentmap.mjs
agentmap: 154 files | 4 features | top hub: lib/utils.ts (deg 52, pr 0.105171)

The --any router

Don't want to learn eight flags? You don't have to. Throw anything at --any — a filename, a function, a feature, even a raw string — and it figures out what you meant, returning the first layer that hits:

--any <query>
   │
   ├─ 1. FILE     exact path → unique basename → unique substring
   ├─ 2. SYMBOL   exported name contains the query (across all files)
   ├─ 3. FEATURE  app/-router feature name contains the query
   └─ 4. CONTENT  live `git grep` (tracked + untracked) — never stale

Layers 1–3 read the cached structural map (fast, ranked). Layer 4 is a live disk read via git grep -F, so raw strings, copy, Tailwind classes, and config values the structural graph never indexes still resolve instead of coming up empty.

Symbol hit (query resolved to a symbol → full block):

$ node agentmap.mjs --any cn
[structure] 1 symbol, 0 feature match for "cn"
  lib/utils.ts → cn (FunctionDeclaration)

Ambiguous file hit (query matched multiple files → narrow it):

$ node agentmap.mjs --any utils
[structure] "utils" matched 3 files — narrow it:
  lib/utils.ts
  lib/db/utils.ts
  tests/prompts/utils.ts

Content fallback (no file/symbol/feature match → live git-grep):

$ node agentmap.mjs --any streamText
[content] 13 lines:
app/(chat)/api/chat/route.ts:8:  streamText,
app/(chat)/api/chat/route.ts:194:        const result = streamText({
artifacts/code/server.ts:1:import { streamText } from "ai";
artifacts/code/server.ts:18:    const { fullStream } = streamText({
artifacts/code/server.ts:40:    const { fullStream } = streamText({
artifacts/sheet/server.ts:1:import { streamText } from "ai";
artifacts/sheet/server.ts:11:    const { fullStream } = streamText({

Commands

Every snippet below is representative output (long lists trimmed) from running agentmap against the public 154-file Next.js repo vercel/ai-chatbot (sha 2becdb4).

--any <q> — the router (file → symbol → feature → live content)

See The --any router above. Default first move for any "where/what/who" question.

--find <q> — reuse-before-rebuild symbol search

Find every exported symbol whose name contains the query. Use it before writing a new util or component to check what already exists.

$ node agentmap.mjs --find Message
find "Message": 55 match
  hooks/use-messages.tsx → useMessages (FunctionDeclaration)
  lib/errors.ts → getMessageByErrorCode (FunctionDeclaration)
  lib/types.ts → messageMetadataSchema (VariableDeclaration)
  lib/types.ts → MessageMetadata (TypeAliasDeclaration)
  lib/types.ts → ChatMessage (TypeAliasDeclaration)
  lib/utils.ts → convertToUIMessages (FunctionDeclaration)
  lib/utils.ts → getTextFromMessage (FunctionDeclaration)
  tests/helpers.ts → generateTestMessage (FunctionDeclaration)
  app/(chat)/actions.ts → generateTitleFromUserMessage (FunctionDeclaration)
  …

--relates <path> — blast radius + transitive relevance

The file's own block (exports / imports / direct dependents) plus a random-walk relevance list (personalized PageRank on the bidirectional import graph) — the files most related to the target, transitively, not just its direct importers.

$ node agentmap.mjs --relates lib/db/schema.ts
relates: lib/db/schema.ts  (pr 0.073744)
exports (14): user(VariableDeclaration), User(TypeAliasDeclaration), chat(VariableDeclaration), Chat(TypeAliasDeclaration), message(VariableDeclaration), DBMessage(TypeAliasDeclaration), …
imports (0): —
dependents (21): hooks/use-active-chat.tsx, lib/types.ts, lib/utils.ts, components/chat/artifact.tsx, components/chat/message.tsx, lib/db/queries.ts, app/(chat)/api/chat/route.ts, …
related (random-walk relevance):
  lib/utils.ts (0.0476)
  lib/types.ts (0.0376)
  components/chat/artifact.tsx (0.0372)
  components/chat/icons.tsx (0.0264)
  components/chat/message.tsx (0.0237)
  lib/db/queries.ts (0.0225)
  app/(chat)/api/chat/route.ts (0.0218)
  …

--feature <name> — files that make up a feature

Resolves a Next.js app/-router feature to its file set, plus the external files that depend on it.

$ node agentmap.mjs --feature api
feature "api": 11 files
  app/(chat)/api/chat/route.ts
  app/(chat)/api/chat/schema.ts
  app/(chat)/api/document/route.ts
  app/(chat)/api/history/route.ts
  app/(chat)/api/messages/route.ts
  app/(chat)/api/models/route.ts
  app/(chat)/api/suggestions/route.ts
  app/(chat)/api/vote/route.ts
  app/(auth)/api/auth/guest/route.ts
  app/(chat)/api/files/upload/route.ts
  app/(chat)/api/chat/[id]/stream/route.ts
external dependents (0): —

--features — list features by size

$ node agentmap.mjs --features
features (4):
  api (11 files)
  login (1 files)
  register (1 files)
  chat (1 files)

--hubs — most important files (PageRank)

The files that matter most, ranked by PageRank importance (raw dependent degree shown alongside).

$ node agentmap.mjs --hubs
agentmap: 154 files (sha 2becdb4)
hubs (PageRank importance):
  lib/utils.ts (deg 52, pr 0.105171)
  lib/db/schema.ts (deg 21, pr 0.073744)
  lib/types.ts (deg 23, pr 0.067589)
  components/chat/artifact.tsx (deg 15, pr 0.036882)
  components/chat/icons.tsx (deg 27, pr 0.035378)
  lib/errors.ts (deg 9, pr 0.032787)
  lib/db/queries.ts (deg 14, pr 0.030085)
  …

--symbols [N] — top ranked symbols (Aider-style)

The most important individual symbols across the repo, ranked by the identifier graph (defaults to 30).

$ node agentmap.mjs --symbols 10
top 10 ranked symbols (Aider-style):
  0.109902  lib/utils.ts → cn (FunctionDeclaration)
  0.036013  lib/types.ts → ChatMessage (TypeAliasDeclaration)
  0.025686  components/chat/artifact.tsx → ArtifactKind (TypeAliasDeclaration)
  0.022461  lib/errors.ts → ChatbotError (ClassDeclaration)
  0.021068  lib/types.ts → CustomUIDataTypes (TypeAliasDeclaration)
  0.020872  lib/db/schema.ts → Document (TypeAliasDeclaration)
  0.020555  components/ai-elements/suggestion.tsx → Suggestion (VariableDeclaration)
  0.020555  lib/db/schema.ts → Suggestion (TypeAliasDeclaration)
  0.018124  lib/db/schema.ts → DBMessage (TypeAliasDeclaration)
  0.015034  lib/errors.ts → ErrorCode (TypeAliasDeclaration)

--map [--tokens N] [--focus <path>] — token-budgeted ranked digest

The token-budgeted digest (Aider's killer feature): a ranked, files-and-symbols summary that fits a token budget. Default budget is 8192 (1024 with --focus). --focus <path> personalizes the ranking toward a file you're working on.

$ node agentmap.mjs --map --tokens 400
# agentmap (154 files, sha 2becdb4) — focus: global, budget ~400 tok

lib/utils.ts:
  cn (FunctionDeclaration)
  generateUUID (FunctionDeclaration)

lib/types.ts:
  ChatMessage (TypeAliasDeclaration)
  CustomUIDataTypes (TypeAliasDeclaration)
  ChatTools (TypeAliasDeclaration)
  Attachment (TypeAliasDeclaration)

components/chat/artifact.tsx:
  ArtifactKind (TypeAliasDeclaration)
  UIArtifact (TypeAliasDeclaration)
  Artifact (VariableDeclaration)

lib/errors.ts:
  ChatbotError (ClassDeclaration)
  ErrorCode (TypeAliasDeclaration)

lib/db/schema.ts:
  Document (TypeAliasDeclaration)
  Suggestion (TypeAliasDeclaration)
  DBMessage (TypeAliasDeclaration)

# ~387 tokens (14 files shown)

Focused on a working file — the ranking re-centers on what lib/db/queries.ts actually touches:

$ node agentmap.mjs --map --focus lib/db/queries.ts --tokens 350
# agentmap (154 files, sha 2becdb4) — focus: lib/db/queries.ts, budget ~350 tok

lib/utils.ts:
  cn (FunctionDeclaration)
  generateUUID (FunctionDeclaration)
  getDocumentTimestampByIndex (FunctionDeclaration)
  fetcher (VariableDeclaration)
  getTextFromMessage (FunctionDeclaration)
  convertToUIMessages (FunctionDeclaration)
  fetchWithErrorHandlers (FunctionDeclaration)
  sanitizeText (FunctionDeclaration)

lib/db/schema.ts:
  DBMessage (TypeAliasDeclaration)
  Suggestion (TypeAliasDeclaration)
  Document (TypeAliasDeclaration)
  Chat (TypeAliasDeclaration)
  User (TypeAliasDeclaration)
  chat (VariableDeclaration)
  document (VariableDeclaration)
  message (VariableDeclaration)

lib/errors.ts:
  ChatbotError (ClassDeclaration)
  ErrorCode (TypeAliasDeclaration)

# ~324 tokens (8 files shown)

--print — full map as JSON

Dumps the cached map (hubs, features, rankedSymbols, files) as one JSON object — for piping into other tools. Also includes a top-level fileCount.

$ node agentmap.mjs --print | jq '.hubs[0]'
"lib/utils.ts (deg 52, pr 0.105171)"

Global flags

| Flag | Description | |------|-------------| | --help / -h | Print a usage block listing every flag and exit 0. | | --version / -v | Print the version from package.json and exit 0. | | --json | Global modifier. When present, every command prints exactly one JSON object to stdout (no prose). Shapes vary per command: --json --hubs → {command,fileCount,sha,hubs:[string]}, --json --find X → {command,query,matches:[{file,name,kind}]}, --json --relates X → {command,file,pagerank,exports,imports,dependents,related}, --json --any X → {command,query,kind,…payload}, etc. Bare --json (no query flag) → {command:"build",fileCount,features,topHub}. | | --install-hooks | Copy hooks/post-commit into .git/hooks/ (chmod 0755), ensure .claude/agentmap.json is in .gitignore, and auto-wire the Claude Code PreToolUse(Grep) nudge into .claude/settings.json (merge-safe + idempotent). Exit 0 on success, stderr + exit 1 on failure. | | --hook-status | Report whether the post-commit hook, PreToolUse nudge, and .gitignore entry are installed (no writes). | | --doctor | Read-only harness health report: git/Claude hook wiring, installed skills + Cursor rule freshness vs package.json version, MCP config entries for OpenCode/Antigravity, and map-cache presence/freshness hints. Always exits 0; suggests fix commands (agentmap --install-hooks, --install-skill, --setup-mcp, agentmap) but never runs them. Combine with --json for a structured report. | | --install-skill | Install skills + always-on docs/hooks per platform (--platform claude\|cursor\|codex\|opencode\|gemini\|antigravity\|copilot\|agents\|all, default all; --project default, or --global; --dry-run preview). | | --mcp | Start agentmap as a stdio MCP server so non-Claude-Code agents (Cursor, Cline, any MCP client) can call every flag as a first-class tool. |

Exit-code contract: 0 = success / match / help / version; 1 = query returned zero results (--any, --find, --relates, --feature with no match); 2 = usage error (missing required arg, unknown flag). Any token starting with - that matches no known flag prints an error to stderr and exits 2.

Scope & limitations

Honesty first — this is deliberately a small, sharp tool, not a universal code-graph.

• TS/JS only, by design. Built on ts-morph. No Python, Go, Rust, etc. If your repo isn't TypeScript/JavaScript, use a tree-sitter-based tool instead. Support for other languages is a possible future direction.
• File-level import graph, not a full reference graph. Edges come from static import / re-export declarations and the named symbols crossing them. It does not do call-site or full reference resolution — --relates tells you which files import a module, not every line that calls a given function.
• PageRank + symbol ranking are real and implemented (damping 0.85, deterministic power iteration; personalized variants for --relates and --map --focus). The symbol ranking is a faithful port of Aider's identifier-graph approach (credit: Aider, Apache-2.0).
• Feature detection assumes the Next.js app/ router. --feature / --features derive features from the first real route segment under app/ (or src/app/), skipping route groups (...), dynamic [...], and parallel @... segments. Repos without an app/ directory simply report zero features — every other command still works.
• Token counts are estimates (chars / 4), not a real BPE tokenizer. Treat --map/--tokens budgets as approximate (±10%).
• The PreToolUse hook is Claude Code-specific (it speaks Claude Code's hook JSON). The post-commit hook is generic git.

Contributing

Issues and PRs welcome. High-value directions:

• Retrieval-accuracy eval — done (EVAL.md, npm run eval). Next: a type-aware dependents mode (the eval excludes type-only edges to match the value-import graph) and an app/-router fixture so --feature retrieval can be scored too.
• A real tokenizer behind the --map budget.
• Hardening feature detection for non-app/-router layouts.

Keep the dependency footprint minimal — ts-morph is the only runtime dependency (it bundles the TypeScript compiler, ~10 MB installed), and keeping it that way is a feature.

License

MIT. Symbol-ranking algorithm credit: Aider (Apache-2.0).