About lens-mcp
lens-mcp is an MCP server for AI agents that maps code and Markdown, including TypeScript, JavaScript, Python, and Prisma structures. It combines project overviews, symbol navigation, reference lookup, and focused reads, while reporting parse errors and output truncation.
Key features
- 01Map code and docs together
- 02Parse code and Prisma structures
- 03Find definitions and symbol references
- 04Read a function or doc section
- 05Search Markdown content
- 06Report errors and truncation totals
Use cases
- 01Get oriented in an unfamiliar project
- 02Locate symbol definitions and usages
- 03Read only a relevant function or doc section
MCP support
It exposes 13 navigation tools to MCP clients, including Claude Code, Cursor, and Codex.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
We haven’t collected a package or remote endpoint for this server. Follow the instructions in the project README.
Tools
Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.
Details
From the project READMElens-mcp
Deterministic navigation maps over code AND markdown — for AI agents. One MCP server, two lenses: tree-sitter for source (TypeScript / JavaScript / Python) and a markdown lens for docs. It answers "where is X and what's the shape of this project?" in one cheap call — so an agent spends context on thinking, not on browsing files.
Speaks the Model Context Protocol; works with any MCP client — Claude Code, Cursor, Codex, or your own agent.
Why
An agent orienting in an unfamiliar repo otherwise burns tokens ls-ing,
grep-ing, and Reading whole files to find the one function or the right
doc. lens returns the map instead of the territory:
function_bodyreads one function — often ~99% less context than Reading the file it lives in.headingreads one doc section — the referenced heading and its subsections, nothing else.mapreturns the whole project's surface — every code file's structure and every doc's outline — in a single call.
Every output is deterministic (real parsing, not a model summarizing), capped
with an honest truncated flag, and framed by one contract:
lens is a navigation map. Use it to locate, then Read the real source/section before judging or modifying it. A signature is not the body; an outline is not the section.
Tools (13)
Orientation
| Tool | What it does |
|---|---|
map | Whole-tree surface in one call: per code file → structure; per doc → title + outline. Both families, one response. |
info | Version, sandbox root, supported languages/extensions, tool list, every output cap, and the lens contract. |
lens_system | Install status, self-update (update.sh under the hood), and fetch the current AGENTS.md — lens's self-maintenance tool. |
Code (tree-sitter — .ts .tsx .mts .cts .js .jsx .mjs .cjs .py) + Prisma (.prisma)
| Tool | What it does |
|---|---|
overview | One file's imports, exports, classes (+ methods), top-level functions, with line ranges. |
functions | Every function incl. nested — signatures, params/types, parent scope, kind. |
function_body | Verbatim source of ONE function — the focused read. |
comments | Comments + TODO/FIXME/BUG/HACK/… markers (markersOnly for the debt list). |
find | Locate a definition by name — functions, classes, and now const/type/enum/exports — across a directory. |
references | The inverse: who uses a symbol — call sites, imports, type-refs — tree-sitter-precise, no grep false positives. |
Docs (markdown — .md .markdown .mdx)
| Tool | What it does |
|---|---|
outline | Full heading hierarchy (the TOC) with line numbers. |
heading | Read ONE section by heading text / slug / line number. |
links | Extract inline / image / wikilink / autolink / reference links. |
search | Case-insensitive full-text search across docs (heading hits ranked first). |
overview/find/map also cover Prisma schemas (schema.prisma → models, enums, fields, relations). JSON config/i18n isn't structurally mapped — lens says so honestly and points you to grep.
Call a code tool on a .md (or a doc tool on a .ts) and it fails with a
helpful pointer to the right tool — no silent confusion.
Honest by construction
- Never silent data loss — a file with syntax errors returns
hasErrors+parseErrors, still extracting what it can; unparseable files inmap/findappear with an inlineerror, never vanish. - Caps everywhere — every list is bounded (see
info.limits) and every cap is reported with the true total. A context-saving tool with unbounded output is self-refuting. - Path sandbox — only files under the server's working directory are
readable; escaping symlinks are rejected.
inforeports the root. - Errors name the fix —
{error, path, hint}, withisErrorset.
Install
One line clones, installs, and writes a ready-to-paste MCP config with this install's absolute path:
git clone https://github.com/segentic-lab/lens-mcp && cd lens-mcp && ./install.sh
No system packages and no native build — tree-sitter runs as WebAssembly, so it
works anywhere Node 18+ runs (Linux, macOS, Windows via WSL/Git Bash). The
installer checks Node, runs npm ci, builds (tsc → dist/), self-tests the full
suite (143 tests), and generates mcp-config.json.
Register it with your client — the generated mcp-config.json looks like:
{
"mcpServers": {
"lens": { "command": "node", "args": ["/abs/path/lens-mcp/dist/index.js"] }
}
}
- Claude Code:
claude mcp add lens -- node /abs/path/lens-mcp/dist/index.js(or copymcp-config.jsoninto a project as.mcp.json) - Cursor / Windsurf: merge
mcp-config.jsoninto~/.cursor/mcp.json - Codex CLI: add
[mcp_servers.lens]with the same command/args to~/.codex/config.toml
lens reads files under its working directory — the project your client
launches it in. Point it at a project and call map("."). Nothing outside the
working directory is readable.
Update later, in place:
./update.sh # git pull --ff-only + reinstall + self-test
Prefer to do it by hand? npm install && npm run build && npm test, then
run node dist/index.js.
Lineage
lens-mcp supersedes the earlier split servers codelens-mcp (code) and
docslens-mcp (docs) — same engines, one server, one pipeline. Sibling of
periscope-mcp (web-app QA);
built to the same standard: honest errors, caps + truncated flags everywhere,
docs == behavior, tests before release.
Built by Segentic Lab. AGPL-3.0.
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.