cocoindex-code-rs

Rust implementation of a local MCP code search server for Claude Code, Codex CLI, and other MCP clients.

This version is pure API mode for embeddings:

• it does not bundle a local embedding model
• it expects a hosted embedding API configured by the user
• the tool/runtime can be free, but embedding API costs depend on the provider

Install

Install the latest release binary with one command:

curl -fsSL https://raw.githubusercontent.com/0xRouteAI/cocoindex-code-rs/main/install.sh | bash

This installs the prebuilt release binary to:

~/.local/bin/cocoindex-code-rs

The installed global CLI command is:

cocoindex-code-rs

Register MCP

Claude Code:

claude mcp add cocoindex-code-rs -- cocoindex-code-rs

Codex CLI:

codex mcp add cocoindex-code-rs -- cocoindex-code-rs

After registration:

• Claude Code or Codex CLI starts a single local MCP process when needed
• the MCP process keeps a project-local index under .cocoindex_code/
• first use auto-builds the project index if missing
• later updates are incremental and only changed files are re-indexed
• while the MCP session stays alive, the server watches accessed projects in the background

Agent Prompt

Add this to your project's AGENTS.md:

Use the `cocoindex-code-rs` MCP server for semantic code search when:
- searching by behavior or meaning instead of exact text
- exploring unfamiliar parts of the codebase
- looking for similar implementations
- grep or filename search is noisy or inconclusive

Prefer normal text search first when exact names, symbols, routes, config keys, or error strings are known.

When using semantic search:
- use it to find candidate files and code chunks
- then verify results with local code reading or text search
- avoid repeated semantic searches when one search already identified the area

What It Does

• Exposes an MCP server over stdio with:
  • search_code
  • index_project
  • project_status
• Builds a local SQLite + sqlite-vec index per project
• Stores project data under each project's .cocoindex_code/
• Uses syntax-aware chunking with tree-sitter when available
• Auto-indexes on first use
• Performs incremental indexing on later searches
• Watches accessed project files in the background during the active MCP session

Settings

User settings file:

~/.cocoindex_code/settings.yml

Example:

api_key: sk-your-api-key
api_base: https://api.openai.com/v1
model: text-embedding-3-small
embedding_dim: 1536

Environment variable equivalents:

OPENAI_API_KEY
OPENAI_API_BASE
EMBEDDING_MODEL
EMBEDDING_DIM

This project does not run local embedding inference by itself.

Build From Source

cargo build --release
./target/release/cocoindex-code-rs

Basic CLI

cocoindex-code-rs init
cocoindex-code-rs index /path/to/project
cocoindex-code-rs search "authentication logic" --project-root /path/to/project
cocoindex-code-rs status --project-root /path/to/project
cocoindex-code-rs mcp --project-root /path/to/project

Notes

• path filtering currently falls back to a full SQL distance scan
• language filtering can use vec partition-aware search
• the MCP server is single-process and does not require a separate local daemon
• each project keeps its own local index database

Benchmarks

Real-world measurements on this repository with a hosted embedding API:

• Whole monorepo, persistent MCP process, 100 natural-language queries:
  • Top1: 39%
  • Top5: 62%
  • median latency: ~322 ms
• Rust subproject only (cocoindex-rs), persistent MCP process, 100 queries:
  • Top1: 59%
  • Top5: 89%
  • median latency: ~349 ms

Practical guidance:

• best results come from searching within the current project or language scope
• whole-monorepo semantic search is usable, but precision is lower across mixed languages and docs
• persistent MCP sessions are much faster than repeated CLI cold starts

More Docs

• INSTALL.md
• MCP_SETUP.md
• USAGE_BILINGUAL.md