io.github.SylphxAI由 GitHub 帳號 SylphxAI 發布v1.9.0更新於 Oct 29, 2026
A map of your codebase for AI agents: code graph, search, call paths and change impact. No API key.
給 AI 助理的本地程式碼庫地圖：程式碼圖、混合檢索、呼叫路徑、變更影響與資料庫結構對應，結果皆標註 file:line。
Desktop only，透過 STDIO。 STDIO 服務會啟動本機處理程序，因此需要 SourceWeft 桌面主機。
參照 儲存庫 中的啟動說明。
Code graph · hybrid search · call paths · change impact · an interactive graph UI.
One Rust binary. Local. No API key. MIT.
[npm] [npm downloads] [CI] [MCP Registry] [License: MIT] [agent-ready 93/100]
Live demo · Docs · Quickstart · Tools · Graph UI · Benchmarks · Compare
[repomap demo: the map of excalidraw, searching restoreElements, its code and callers, then the impact of changing it]The real UI on excalidraw (687 files, indexed in under half a second): search, a symbol's code and callers, then the blast radius of a change. Try it in your browser, no install needed.
Run it on any repository. This is the real output on repomap's own source (v1.5.0, trimmed): ask what breaks if pagerank changes.
One local call, no API key, and an answer your agent can cite as file:line.
Prefer to wire one client by hand? claude mcp add repomap -- npx -y @sylphx/repomap mcp for Claude Code; the Cursor, VS Code, Codex and Docker snippets follow.
That's it. Add --claude-hooks to also enrich Claude Code's Grep and Glob. setup detects the clients you have, writes their MCP config, and prints every change it made. Run it again and nothing changes. Then ask your agent: "Use repomap to map this repo."
[Add to Cursor] · Install in VS Code
Requires Node 18+. Confirm the install in your editor, open your project, and enable the server. Both links run npx -y @sylphx/repomap mcp locally; no API key is needed.
For VS Code from a terminal:
Or add the server to .cursor/mcp.json:
For .vscode/mcp.json, use the VS Code servers format:
Claude Code: claude mcp add repomap -- npx -y @sylphx/repomap mcp
Claude Desktop, one click: download repomap-<version>.mcpb from the latest release, open it, and pick your project folder.
Claude Code plugin: /plugin marketplace add SylphxAI/repomap, then /plugin install repomap@repomap
Codex (~/.codex/config.toml):
Docker (stdio, amd64/arm64):
The server indexes the client's workspace root (or its working directory, or REPOMAP_ROOT). Every tool also takes root.
One issue-text search finds every fix file in the top 10 for 83.0% of SWE-bench Verified issues (semble: 71.0%). File retrieval, not issue resolution.
[SWE-bench Verified: repomap finds every fix file in the top 10 for 83.0% of 500 issues; semble 71.0%, BM25 55.8%. Full Acc@1, 5 and 10 values follow in the table.]Defaults selected on 300 disjoint SWE-bench test instances, excluding every Verified ID, then frozen before one evaluation on all 500 Verified instances. One search per issue, at its base commit; gold files are used only for scoring. All three methods scored 500/500, with zero errors.
Split definitions, pinned revisions and full results · Committed chart data
Trade-off: on semble's public code-search set, NDCG@10 is 0.845 versus semble's 0.851; semble wins there. This is not a claim of better search on every task. Measured search quality and limitations.
Agents burn most of their context on grep, ls and reading whole files just to work out where things are. repomap gives them the map up front:
parseConfig.file:line.git diff before you commit.All of it comes from a local index: tree-sitter parsing, a resolved import and call graph, PageRank, Louvain communities, BM25 over AST chunks, and a small static code embedding model (33 MB, downloaded once from Hugging Face, then offline). Nothing leaves your machine, and no API is called; the model download is the only network request (set REPOMAP_EMBED=0 to skip it).
Six tools, each with an obvious job:
Answers are compact text that cites file:line, so they cost few tokens. Pass format: "json" for structured output.
The same commands work in your terminal: repomap map, repomap search "…", repomap context X, repomap trace A B, repomap impact --changed, repomap db.
export writes one HTML file with deep links (#path/to/file) and GitHub links pinned to your commit. It's a good fit for a README, a wiki or a design review.repomap reads your schema from the repository: SQL migrations (applied in order, with down migrations skipped), schema.prisma, Drizzle pgTable/mysqlTable/sqliteTable, SQLAlchemy and Flask-SQLAlchemy models, Diesel table!, and Django models.py (fields, ForeignKey/ManyToManyField, Meta, implicit join tables). It can also introspect a live database. Each table is linked to the code that queries it: raw SQL (FROM users), Prisma (prisma.user.findMany), Diesel (users::table), Django (Post.objects…), and ORM models or tables used by files that import them.
Live connections are strictly read-only:
READ ONLY transaction that the server must confirm.START TRANSACTION READ ONLY transaction.query_only.Only catalog metadata is read, never table rows. The connection string comes from an argument or an environment variable, and it is never stored or printed. Agents get the same data through the db MCP tool; pass url_env rather than the URL itself.
The score rates how well an AI agent can work in the repository, from 0 to 100. It covers eight checks:
Each check that falls short comes with a concrete fix, and the score ends with a badge line for your README. For CI, use --min 70; --update-readme README.md refreshes the badge.
Keep the badge fresh with the GitHub Action:
This is opt-in and safe to run again; setup --remove takes it out. It installs a PreToolUse hook: whenever Claude Code runs Grep or Glob, repomap adds where the symbol is defined, who calls it and which module it belongs to. It answers in tens of milliseconds and never blocks the search.
Parsed with tree-sitter for symbols, calls, imports and inheritance: TypeScript, TSX, JavaScript, Python, Go, Rust, Java, Kotlin, Swift, C, C++, C#, Ruby, PHP. Search also covers Markdown, YAML, TOML, JSON, SQL, shell, Protobuf, GraphQL, HTML/CSS, Vue, Svelte, Scala and more.
Modules are found among your core code only. Tests, examples, docs and benchmarks are grouped separately, so they never name or blur a module.
Import resolution understands relative paths, @/ aliases, npm workspace packages, Python packages and relative imports, Go modules, Rust mod/use/workspace crates, Java/PHP namespaces, C/C++ includes and Ruby require. .gitignore is respected, and so is .repomapignore.
On semble's public code-search benchmark (63 repositories, 19 languages, 1,251 questions), repomap's search scores NDCG@10 0.845, versus semble 0.851 on the same runner. The previous repomap score was 0.851: stronger issue localization trades about 0.006 here. On our 60 large-repository questions, repomap scores 0.846 versus semble 0.799, but semble still wins on VS Code. The 137M-parameter CodeRankEmbed model's published score is 0.839 and plain BM25's 0.673 (cited, not re-run).
On file localization (500 SWE-bench Verified issues, one search call with the issue text), repomap finds every fix file in the top 10 for 83.0% of issues (Acc@10; 74.8% at 5, 48.8% at 1), versus semble's 71.0% / 61.6% / 34.0%. Before this change, repomap scored 62.2% / 51.8% / 23.6%. Plain BM25 scores 55.8% at 10, and repomap with embeddings off 77.6%. Same-run median index/query costs are 1.682 s / 54.4 ms, versus the old binary's 1.686 s / 59.8 ms and semble's 12.813 s / 354.4 ms. Defaults were selected only on a disjoint 300-instance tuning split, then frozen before one Verified measurement. Agent systems such as LocAgent report higher numbers on a different subset with a language model in the loop; those are cited, not re-run.
This needs no GPU, no vector database and no API key. A 33 MB static code model runs on the CPU, next to BM25 and symbol names.
The index is built in parallel and cached per file, so after the first run only changed files are parsed again. The MCP server keeps the graph in memory and refreshes it when files change.
Measured on a 4 vCPU GitHub-hosted runner (method and full table):
Pick Serena if you want LSP-precise refactoring edits. repomap is for understanding and navigating a codebase with zero setup, including semantic search without a vector database or an API key, and a licence you can use at work. Search quality is measured on public benchmarks in Benchmarks.
Everything repomap does is free under MIT, including the multi-repository workspace (--workspace) and impact in your own CI. There is no licence key and no paid tier.
Binaries for macOS (arm64, x64), Linux glibc (x64, arm64) and Windows x64 ship as npm optional dependencies. They're also attached to each GitHub release. From source: cargo install --git https://github.com/SylphxAI/repomap sylphx-repomap.
@sylphx/spine, @sylphx/locus and @sylphx/coderag are thin aliases of @sylphx/repomap, and their old tool names still work. See the migration guide for the name mapping and how to switch.
Issues and PRs are welcome, and new language support is especially useful: add a grammar and a query in crates/repomap-core/src/lang.rs.
More from Sylphx: https://sylphx.com/open-source
repomap runs on your machine. Indexing, search and every query read your code
locally and send nothing anywhere. The one network request repomap makes is a
single download of the embedding model from Hugging Face the first time you
search; it sends no code. Set REPOMAP_EMBED=0 to skip it and use keyword
search only. repomap export files hold paths, symbol names and line numbers,
never source. See SECURITY.md for the full boundary.
Questions, a bug, or a team rollout? Email [email protected] or open an issue.
MIT licence. © Sylphx Limited, registered in England and Wales, company no. 16438428. Registered office: 128 City Road, London EC1V 2NX, United Kingdom. Email [email protected]. Privacy · Terms · sylphx.com
來源：README.md，提交 b5c353e
Map the codebase
Start here. A map of the repository: modules (communities of files that depend on each other), the most central files (PageRank), the most used symbols, and entry points. Pass `focus` (a directory) to zoom in and get an outline of its files and symbols with line numbers.
rootfocuslimitformattokensSearch code
Hybrid code search over AST chunks (whole functions, methods, classes): symbol names, BM25 keywords and a local code embedding model, so identifiers and plain questions both work, e.g. "parseConfig", "refresh token expiry" or "where are failed requests retried". Returns ranked file:line ranges with the matching lines.
kindpathrootlimitquery必填formatinclude_tests360° view of a symbol or file
Everything about one symbol or file: its code, who calls it (with call-site lines), what it calls, subtypes, members, imports, importers and the tests that touch it. Large fixture trees are skipped until targeted (target one, or pass --include-fixtures). Target forms: `path/to/file.ts`, `file.ts:42`, `Class.method`, `Class::method`, or a bare name.
rootformattarget必填code_linesTrace call paths
With `from` and `to`: the shortest call path between two symbols or files, each hop cited file:line (falls back to the file dependency path). With only `from`: the call tree below it (direction callees) or above it (direction callers).
tofrom必填rootdepthformatdirectionChange impact (blast radius)
What breaks if this changes. Give `target` (symbol or file, or a list) or `changed: true` to analyse the current git diff. Returns a risk level, direct and indirect callers with call sites, importing files, modules touched, and the tests to run. Fixture files in deferred trees are not analysed unless targeted; the result says so.
baserootdepthformattargetchangedDatabase map
Map the database: tables, columns, primary and foreign keys, indexes, and the code that queries each table (file:line). By default it reads schema sources in the repo: SQL migrations, Prisma, Drizzle, SQLAlchemy, Diesel and Django models. To inspect a live Postgres, MySQL or SQLite database, pass `url_env` (the NAME of an environment variable holding the connection string). The connection is strictly read-only and the string is never stored or shown. Pass `table` for one table's full detail.
urlroottableformaturl_env已滿足適用檢查項的 90%
該評分衡量此收錄條目的完整程度，依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價，也不表示伺服器安全或目前可用：可連通性是帶日期的檢查結果，評分也不包含安裝量或使用者評價。
請在註冊中心條目中宣告憑證請求標頭，或保持端點能夠回應探測，以便判定其認證方式。
npx -y @sylphx/repomap setup$ npx -y @sylphx/repomap impact pagerank
# Impact (MEDIUM risk)Changing: function pagerank (crates/repomap-core/src/graph.rs:639)2 direct callers, 4 symbols affected in total across 6 files and 3 modules; 5 files importing the changed files; 1 test file to run.
## Direct callers (will break if the contract changes)- method DbSchema.graph_json — crates/repomap-core/src/db.rs:1981- function link — crates/repomap-core/src/graph.rs:122
## Indirect (depth 2)- function assemble — crates/repomap-core/src/index.rs:700npx -y @sylphx/repomap setup # add repomap to Claude Code, Codex, Cursor, VS Code, Claude Desktop, Windsurf, Gemini CLInpx -y @sylphx/repomap serve # open the graph UI for the current repocode --add-mcp '{"name":"repomap","command":"npx","args":["-y","@sylphx/repomap","mcp"]}'{ "mcpServers": { "repomap": { "command": "npx", "args": [ "-y", "@sylphx/repomap", "mcp" ] } }}{ "servers": { "repomap": { "type": "stdio", "command": "npx", "args": [ "-y", "@sylphx/repomap", "mcp" ] } }}{ "mcpServers": { "repomap": { "command": "npx", "args": ["-y", "@sylphx/repomap", "mcp"] } }}[mcp_servers.repomap]command = "npx"args = ["-y", "@sylphx/repomap", "mcp"]{ "mcpServers": { "repomap": { "command": "docker", "args": ["run", "-i", "--rm", "-v", "/path/to/repo:/workspace:ro", "ghcr.io/sylphxai/repomap"] } } }> impact target=decode
# Impact (LOW risk)Changing: function decode (src/auth/token.ts:14)1 direct caller, 3 symbols affected in total across 3 files and 2 modules; 1 file importing the changed files; no tests reach this.
## Direct callers (will break if the contract changes)- function verifyToken — src/auth/token.ts:8
## Indirect (depth 2)- method SessionStore.refresh — src/auth/session.ts:5
## Indirect (depth 3)- function handleRefresh — src/api/router.ts:6
## Importing files- src/auth/session.tsnpx -y @sylphx/repomap serve # live, with code previewnpx -y @sylphx/repomap export # repomap.html: one self-contained file you can publishnpx -y @sylphx/repomap db # from migrations / Prisma / Drizzle / SQLAlchemy / Diesel / Djangonpx -y @sylphx/repomap db --url-env DATABASE_URL # live Postgres, MySQL or SQLite, read-onlynpx -y @sylphx/repomap db users # one table: columns, indexes, who references it, where it is queriednpx -y @sylphx/repomap db --serve # the same graph UI, for tables and foreign keysnpx -y @sylphx/repomap score# .github/workflows/agent-ready.ymlon: { push: { branches: [main] } }permissions: { contents: write, pull-requests: write }jobs: score: runs-on: ubuntu-latest steps: - uses: actions/checkout@v4 - uses: SylphxAI/repomap@v1 with: update-readme: true # opens or updates one PR with the fresh badge min-score: 0 # set e.g. 70 to fail below a barnpx -y @sylphx/repomap setup --claude-hooksrepomap (code map) for this search:- function `compose` defined at src/compose.ts:15 (module router); 3 callers: Hono.route (src/hono-base.ts:228), Hono.#dispatch (src/hono-base.ts:452), every (src/middleware/combine/index.ts:102)repomap setup [--client cursor,codex] [--claude-hooks] [--dry-run] [--remove]repomap serve [dir] [--port 7878] [--no-open]repomap export [dir] [--out repomap.html] [--json]repomap map [dir-to-focus] [-C root] [--tokens 1000] [--json]repomap search <query> [--path src/] [--kind function] [--limit 10]repomap context <target> [--code-lines 60]repomap trace <from> [to] [--callers] [--depth 3]repomap impact [targets…] [--changed] [--base main]repomap db [table] [--url-env VAR | --url URL] [--serve | --out db.html] [--json]repomap score [dir] [--json] [--min N] [--update-readme README.md [--insert]]repomap index [dir] [--no-cache] [--json]repomap mcp [--root dir]cargo test --workspace # engine + CLI testsbun install && bun run build:ui # rebuild the UI bundle (ui/ -> crates/repomap/assets/)cargo run -p sylphx-repomap -- serve .