# codeseeker [Health: Active]

**Category:** 💻 Developer Tools  
**Repository:** https://github.com/jghiringhelli/codeseeker  
**GitHub Stars:** 20  
**npm Downloads (last month):** 640  
**Views:** 1  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/codeseeker

## Description
Graph-powered code intelligence with semantic search and knowledge graph for AI assistants

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "codeseeker": {
    "command": "npx",
    "args": ["-y","work"],
    "env": {
      "CODESEEKER_STORAGE_MODE": ""
    }
  }
}
```

**Requires environment variables:** `CODESEEKER_STORAGE_MODE` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What codeseeker MCP server does

The codeseeker MCP server gives AI coding assistants an indexed view of a repository. It is designed for questions that require more than matching words in files, such as locating authentication logic, finding related implementation patterns, or tracing dependencies across files.

The server builds a knowledge graph from AST-parsed imports and combines that graph with several retrieval methods. Search results can include source files, test files, symbols, and directory-summary nodes. The system also detects coding standards during analysis, allowing assistants to inspect patterns used by a project.

## How it works

Indexing happens before normal querying. During a query, CodeSeeker combines BM25 results with vector search using 384-dimensional Xenova embeddings. Reciprocal Rank Fusion produces an initial set of up to 30 results, including directory summaries generated through the RAPTOR approach.

When a directory summary scores at least 0.5, the retrieval pipeline narrows results to that directory. Otherwise, the full result set continues through the pipeline. Results are then deduplicated so each file keeps its highest-scoring chunk. Additional scoring favors source files, reduces test-file dominance, boosts filename symbol matches, and rewards files with multiple matching chunks.

The highest-ranked results undergo graph expansion. The server follows IMPORTS, CALLS, and EXTENDS relationships from the top 10 results, adding structural neighbors at a reduced score. This helps surface related files that do not contain the query terms directly.

## Setup and configuration

Install the package globally with npm, then register the server with an MCP client:

```bash
npm install -g codeseeker
claude mcp add codeseeker --scope user -e CODESEEKER_STORAGE_MODE=embedded -- codeseeker serve --mcp
```

The `--scope user` option makes the registration available across projects opened by Claude Code. A no-install option uses `npx`:

```json
{
  "mcpServers": {
    "codeseeker": {
      "command": "npx",
      "args": ["-y", "codeseeker", "serve", "--mcp"],
      "env": { "CODESEEKER_STORAGE_MODE": "embedded" }
    }
  }
}
```

Other supported setup paths include `codeseeker install --vscode`, `--cursor`, or `--windsurf`. The Claude Code plugin adds synchronization hooks and the `/codeseeker:init` and `/codeseeker:reindex` commands. Devcontainers and GitHub Codespaces can install the package during container creation.

## Tools and capabilities

The codeseeker MCP server presents one MCP tool with an `action` routing field. Its documented actions are:

- `search` for hybrid semantic and lexical retrieval with graph expansion.
- `sym` for symbol-oriented lookup.
- `graph` for navigating relationships in the code knowledge graph.
- `analyze` for examining project patterns and coding standards.
- `index` for creating or updating the indexed representation.

This single-tool design keeps the per-request MCP tool overhead lower than exposing each action as a separate endpoint.

## Limitations and notes

The initial index is required and can take roughly 30 seconds to five minutes. Global installation avoids making the first MCP startup responsible for downloading and building dependencies; an uncached `npx` launch can be slower.

CodeSeeker addresses retrieval rather than specification, testing, or architecture governance. The repository is licensed under Apache 2.0.

_Full upstream README: https://allmcps.com/mcp/codeseeker/readme_