Introduction
NekoSearch is a local, read-only MCP server for searching documentation snapshots. It gives AI coding agents structured access to the documentation you choose, with explicit source selectors, section-level retrieval, and source revisions attached to the results.
Build a snapshot from your own Markdown or MDX, install a curated source, or ingest a documentation directory from a public GitHub repository. Your coding client launches NekoSearch over stdio and searches the installed snapshots locally.
Why NekoSearch?
An agent needs documentation for the source and version you are working with. NekoSearch makes that choice explicit: discover the available sources, search selected snapshots, and fetch the relevant sections.
- MCP-native retrieval: three focused tools to discover, search, and fetch documentation.
- Your documentation: ingest local Markdown and MDX, public GitHub directories, or portable snapshots.
- Curated sources: install optional NestJS 11 or 12 documentation from the official repository.
- Local snapshots: serve installed documentation without fetching pages or opening a network listener.
- Source-scoped search: search up to five exact selectors together while keeping results grouped by source.
- Internal references: follow resolved document and section links within a snapshot.
- MCP client integration: automatically configure Grok Build CLI, or launch the stdio server from another MCP client using that client's settings.
A small MCP protocol summary is bundled with the CLI. NestJS is optional; no NestJS snapshot is bundled or required.
How it works
Documentation moves through a preparation step before your agent retrieves it:
Documentation sources
→ Ingestion and processing
→ Documentation snapshots
→ Retrieval engine
→ MCP server
→ AI coding agentsEach snapshot contains a normalized corpus, a SQLite search index, and a manifest describing its identity and provenance. The server opens validated indexes read-only when it starts.
The agent calls list_docs to discover available sources, search_docs to search exact selectors, and fetch_docs to read a document or section. Results include canonical URLs and source revisions. See the MCP overview for the complete workflow.
Explicit sources
NekoSearch searches only the sources requested by the client. An unknown selector is an error; it never silently expands a request to the entire catalog.
Get started
Install the CLI with a supported Node.js version:
npm install --global nekosearch
nekosearch setup
nekosearch listInteractive setup lets you select optional curated documentation and review the Grok Build registration when its CLI is available. Other stdio MCP clients can launch the server using their own configuration settings.
Follow the Quick Start to make your first search using the bundled reference, or read Installation for requirements and installation details.
Bring your own documentation
From a project containing Markdown or MDX in a docs directory:
nekosearch add --dir ./docs --id my-api@1Restart the client's NekoSearch process after installation, then ask the agent to discover and search my-api@1. See Custom Documentation for a complete example with internal references.
For other ingestion options, read Documentation Sources. To understand storage, provenance, and updates, continue with Snapshots.
Explore the documentation
- Getting Started: installation, your first search, and interactive setup.
- Guides: documentation sources, snapshots, custom documentation, and updates.
- CLI Reference: supported commands, arguments, options, and errors.
- MCP Tool Reference: tool inputs, structured output, pagination, and references.
- Architecture: the boundaries between the CLI, indexer, engine, catalog, and transport.
- Development: contributor setup and validation.
