Octocode

Code search and dependency graphs for AI assistants.

Quick start · MCP integration · OctoFS · Documentation · Contributing · Website

Octocode is a Rust CLI and MCP server for exploring codebases. It combines semantic search, structural search, and a live graph of symbols and dependencies. Connect it to Claude Code, Codex, Cursor, Windsurf, or another MCP client, or search directly from your terminal.

• Find code by intent with hybrid vector and keyword search, plus configurable reranking.

• Locate definitions and usages with structural search over the current source.

• Trace relationships between files and symbols: imports, calls, inheritance, and implementations.

• Read signatures without loading entire files; add a language server for definitions, references, and hover information.

• Run locally with the default embedding and reranking models, or configure remote providers.

Quick start

1. Install

Use the installer on Linux or macOS, or in a Unix-compatible shell on Windows:

curl -fsSL https://raw.githubusercontent.com/Muvon/octocode/master/install.sh | sh

Alternatively, on macOS:

brew install muvon/tap/octocode

Download a binary from GitHub Releases, or install the published crate:

cargo install octocode

Building from source requires Rust 1.95 or later and protoc. See the installation guide for platform requirements, build features, and manual installation.

Verify the installation:

octocode --version

2. Index and search

Run these commands in your project's Git repository:

cd /path/to/your/project
octocode index
octocode search "authentication middleware"

The default configuration uses local FastEmbed models for embeddings and reranking. No API key is required; model files download on first use. For a directory without Git, use octocode index --no-git.

# Combine related queries
octocode search "auth" "middleware" "session"

# Filter code results by language
octocode search "database connection pool" --language rust

# Search documentation
octocode search "deployment configuration" --mode docs

3. Connect your AI assistant

Choose your client in MCP integration. Structural search, signatures, and the live dependency graph read the current source directly and work without an index. Semantic search uses the index created above.

Related MCP server: codeembed

MCP integration

Install Octocode where your client can find it on PATH, and replace /path/to/your/project with your project's absolute directory path.

Claude Code

claude mcp add --transport stdio octocode -- octocode mcp --path /path/to/your/project
claude mcp list

See the Claude Code MCP documentation for configuration scopes and connection details.

Codex CLI

codex mcp add octocode -- octocode mcp --path /path/to/your/project
codex mcp list

See the Codex MCP documentation for CLI and configuration-file options.

Claude Desktop, Cursor, and Windsurf

Add an Octocode entry to your client's MCP configuration:

{
  "mcpServers": {
    "octocode": {
      "command": "octocode",
      "args": ["mcp", "--path", "/path/to/your/project"]
    }
  }
}

Restart the client after saving. The MCP client setup guide covers configuration locations and other clients.

Octomind

Octomind is a companion AI agent runtime. Its developer:general agent includes Octocode and OctoFS capabilities. After installing Octomind, start it in your project:

cd /path/to/your/project
octomind run developer:general

The first run may install required tools and ask for model credentials. See the Octomind quick start for installation and model access.

Available tools

To enable LSP tools, install a language server and add --with-lsp to the server command, for example:

octocode mcp --path /path/to/your/project --with-lsp rust-analyzer

By default, the MCP server uses the existing semantic index without updating it. Rerun octocode index after changes, or run octocode watch separately in the project directory. In-process indexing is available through index.mcp_index = true.

See the MCP integration guide for tool parameters, HTTP transport, and serving multiple repositories, and the LSP guide for language server setup.

Use Octocode with OctoFS

OctoFS adds file reading and atomic editing tools to your assistant. Run both MCP servers against the same project directory.

Octocode's structural_search returns source lines as N:hh|content, with the same line IDs used by OctoFS view. An agent can use these IDs directly in OctoFS batch_edit, without reading the same lines again to obtain edit targets. Exact-name lookups include complete definitions up to 32 lines; larger results identify omitted ranges for a targeted read.

Each ID hashes the original line, including indentation. OctoFS checks IDs before applying edits and reports stale targets when the content no longer matches. Group known edits to a file into one atomic batch, then use the fresh IDs returned by the edit for follow-up changes. The N:hh| prefix is metadata, not part of the source code.

After installing OctoFS, add it alongside the Octocode server configured above. Choose your client:

Claude Code:

claude mcp add --transport stdio octofs -- octofs mcp --path /path/to/your/project

Codex CLI:

codex mcp add octofs -- octofs mcp --path /path/to/your/project

Use the same absolute project path for both servers. OctoFS is optional; Octocode's search and navigation tools work independently.

How it works

Current source → Language parsers → Live file and symbol graph → MCP tools
Indexed code   → Embeddings + BM25 → Hybrid search + reranking → CLI / MCP

The live graph uses Tree-sitter to extract symbols and relationships from current source, without embeddings or an LLM. It remains available through MCP when [graphrag].enabled = false.

Enabling indexed GraphRAG adds persisted file-level enrichment, including semantic file discovery and optional LLM descriptions and architectural relationships. The octocode graphrag CLI uses this persisted layer:

octocode config --graphrag-enabled true
octocode index
octocode graphrag overview

See the architecture guide for module boundaries and storage details.

Supported languages

Retrieval benchmark

The repository includes ground truth, scoring scripts, and reproduction instructions. The published local matrix used 127 code-search queries, corpus commit b1771ba, Octocode 0.17.1, and fastembed:jinaai/jina-embeddings-v2-base-code, with reranking disabled for the variants below.

These are historical results from June 21, 2026; they do not measure the current release or its default configuration. See the full variant matrix for settings and reranker comparisons.

Privacy and configuration

• Local models by default. Embedding and reranking run locally after model downloads. Remote embedding, reranking, and LLM providers are optional; when enabled, they receive the queries and source content needed for those features.

• Local storage. Indexes live outside your repository in Octocode's per-project data directory.

• File exclusions. Indexing respects .gitignore and .noindex. Configure exclusions for files you do not want indexed.

• Client access. The default MCP transport is stdio. The model used by your AI client may receive source content returned by tools, even when Octocode uses local models.

• Provider credentials. API keys can be supplied through environment variables. See the API key guide before enabling remote providers.

The configuration guide explains model selection and indexing options. The embedded default template contains the current defaults.

Documentation

• Getting started — basic workflow

• Installation — binaries, source builds, and platform requirements

• MCP clients — client setup and configuration locations

• MCP integration — tools, transports, and multiple repositories

• LSP integration — language server tools

• Commands — CLI reference

• Configuration — models and customization

• API keys — remote provider setup

• Architecture — internals

• Changelog — release history

Contributing

Bug reports, documentation improvements, and code contributions are welcome. Read the contributing guide for development setup and the engineering standards before changing code.

Use GitHub Issues for bugs and feature requests, Discussions for questions, or contact opensource@muvon.io.

License

Apache License 2.0. Maintained by Muvon.