AST-based semantic code search that just works

A lightweight, effective (AST-based) semantic code search tool for your codebase. Built on CocoIndex — a Rust-based ultra performant data transformation engine. Use it from the CLI, or integrate with Claude, Codex, Cursor — any coding agent — via Skill or MCP.

• Instant token saving by 70%.
• 1 min setup — install and go, zero config needed!

Get Started — zero config, let's go!

Install

Using pipx:

pipx install 'cocoindex-code[full]'          # batteries included (local embeddings)
pipx upgrade cocoindex-code                  # upgrade

Using uv:

uv tool install --upgrade 'cocoindex-code[full]'

Two install styles — they mirror the Docker image variants of the same names:

• cocoindex-code[full] — batteries-included. Pulls in sentence-transformers so local embeddings (no API key required) work out of the box. The ccc init interactive prompt defaults to Snowflake/snowflake-arctic-embed-xs.
• cocoindex-code (slim) — LiteLLM-only; requires a cloud embedding provider and API key. Use when you don't want the local-embedding deps (~1 GB of torch + transformers).

Next, set up your coding agent integration — or jump to Manual CLI Usage if you prefer direct control.

Coding Agent Integration

This repository is a single plugin marketplace (.claude-plugin/marketplace.json) consumed by both Claude Code and Grok — same plugin id cocoindex-code, same ccc skill. Grok optionally activates the bundled hooks and MCP server with --trust; Claude Code users can install the same marketplace and rely on the skill alone or load hooks/MCP from the plugin as needed.

Skill (Recommended)

Install the ccc skill so your coding agent automatically uses semantic search when needed:

npx skills add cocoindex-io/cocoindex-code

That's it — no ccc init or ccc index needed. The skill teaches the agent to handle initialization, indexing, and searching on its own. It will automatically keep the index up to date as you work.

The agent uses semantic search automatically when it would be helpful. You can also nudge it explicitly — just ask it to search the codebase, e.g. "find how user sessions are managed", or type /ccc to invoke the skill directly.

Works with Claude Code and other skill-compatible agents.

Claude Code plugin marketplace

For Claude Code users, this repository is also a plugin marketplace. Install the skill from inside Claude Code with:

/plugin marketplace add cocoindex-io/cocoindex-code
/plugin install cocoindex-code@cocoindex-code

This bundles the same ccc skill, with version pinning and /plugin marketplace update for updates. The repository also ships hooks/hooks.json and .mcp.json for Grok (and Claude Code plugin installs that load those files); Claude users who want skill-only search can rely on the skill alone and add MCP manually in the MCP Server section below instead of using the bundled .mcp.json.

Grok plugin

For Grok users, install via Grok's plugin system. The plugin bundles three components:

Grok does not import Claude's enabledPlugins or plugin cache; install separately even if you already use cocoindex in Claude Code.

Full install (skill + hook + MCP):

grok plugin marketplace add cocoindex-io/cocoindex-code
grok plugin install cocoindex-io/cocoindex-code --trust
grok plugin enable cocoindex-code

Prefer the GitHub shorthand (cocoindex-io/cocoindex-code) for install — grok plugin install cocoindex-code can fail when no marketplace plugin matches that bare name.

--trust is required so Grok activates the plugin's hooks and MCP server (skills load when the plugin is enabled).

Skill-only (match Claude Code — no auto-index hook, no MCP tool):

Install and enable as above, then disable the optional components:

1. Hooks — open /hooks, select the SessionStart hook from cocoindex-code, press Space to disable.
2. MCP — open /mcps, select cocoindex-code, press Space to disable; or persist in ~/.grok/config.toml:

[mcp_servers.cocoindex-code]
enabled = false

The agent still owns indexing via the ccc skill (ccc index / ccc search --refresh when stale), same as Claude Code.

To avoid importing MCP servers from your Claude/Cursor user config (unrelated to this plugin):

[compat.claude]
mcps = false

[compat.cursor]
mcps = false

MCP Server

Alternatively, use ccc mcp to run as an MCP server:

claude mcp add cocoindex-code -- ccc mcp

codex mcp add cocoindex-code -- ccc mcp

opencode mcp add

Enter MCP server name: cocoindex-code Select MCP server type: local Enter command to run: ccc mcp

Or use opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "cocoindex-code": {
      "type": "local",
      "command": [
        "ccc", "mcp"
      ]
    }
  }
}

Add a local MCP server in ~/.config/kilo/kilo.jsonc, kilo.jsonc, or .kilo/kilo.jsonc:

{
  "mcp": {
    "cocoindex-code": {
      "type": "local",
      "command": ["ccc", "mcp"],
      "enabled": true
    }
  }
}

Once configured, the agent automatically decides when semantic code search is helpful — finding code by description, exploring unfamiliar codebases, fuzzy/conceptual matches, or locating implementations without knowing exact names.

Note: The cocoindex-code command (without subcommand) still works as an MCP server for backward compatibility. It auto-creates settings from environment variables on first run.

When running as an MCP server (ccc mcp), the following tool is exposed:

search — Search the codebase using semantic similarity.

search(
    query: str,                          # Natural language query or code snippet
    limit: int = 5,                      # Maximum results (1-100)
    offset: int = 0,                     # Pagination offset
    refresh_index: bool = True,          # Refresh index before querying
    languages: list[str] | None = None,  # Filter by language (e.g. ["python", "typescript"])
    paths: list[str] | None = None,      # Filter by path glob (e.g. ["src/utils/*"])
)

Returns matching code chunks with file path, language, code content, line numbers, and similarity score.

Manual CLI Usage

You can also use the CLI directly — useful for manual control, running indexing after changing settings, checking status, or searching outside an agent.

ccc init                                # initialize project (creates settings)
ccc index                               # build the index
ccc search "authentication logic"       # search!

The background daemon starts automatically on first use.

Tip: ccc index auto-initializes if you haven't run ccc init yet, so you can skip straight to indexing.

CLI Reference

Search Options

ccc search database schema                           # basic search
ccc search --lang python --lang markdown schema      # filter by language
ccc search --path 'src/utils/*' query handler        # filter by path
ccc search --offset 10 --limit 5 database schema     # pagination
ccc search --refresh database schema                 # update index first, then search

By default, ccc search scopes results to your current working directory (relative to the project root). Use --path to override.

Structural Search (ccc grep)

ccc grep finds code by structure, not text — you write a by-example pattern and it matches the syntax tree (via cocoindex's code_match), so formatting, whitespace, and intervening tokens don't matter. It runs entirely locally: no index, daemon, or embeddings required.

ccc grep 'def \NAME(\(ARGS*\)):'                      # every Python function def under the cwd
ccc grep 'foo(\(ARGS*\))' src/                        # calls to foo(...) anywhere under src/
ccc grep 'fn \NAME(\(A*\))' --lang rust               # restrict to one language
ccc grep 'class \NAME:' --path 'tests/**'            # restrict to a path glob
ccc grep 'TODO(\(A*\))' path/to/file.py               # a single file

Metavariables use the \ sigil: \NAME captures one node, \(NAME*\) a run of siblings, \_/\* match anonymously. The pattern is matched per language, so a single invocation scans every supported source file (others are skipped). Inside an initialized project, ccc grep honors the project's include/exclude patterns and .gitignore; otherwise it scans all supported source files under the path.

Results stream to the terminal file-by-file as each match is found (in completion order, since files are matched in parallel) rather than all at once at the end. Each matching file shows its matched line range; under a TTY the path is colored, line numbers are dimmed, and the unmatched context around a match is dimmed so the match stands out.

Note: ccc grep relies on cocoindex's structural code_match feature. Until it ships in a released cocoindex, run against a local cocoindex build.

Docker

A Docker image is available for teams who want a reproducible, dependency-free setup — no Python, uv, or system dependencies required on the host.

The recommended approach is a persistent container: start it once, and use docker exec to run CLI commands or connect MCP sessions to it. The daemon inside stays warm across sessions, so the embedding model is loaded only once.

Choosing an image

Two variants are published from each release:

The rest of this section uses :latest — substitute :full in the image: / docker run commands if you want the full variant.

Mac users running the :full variant: local embedding inference is CPU-only inside Docker, because Docker on macOS can't access Apple's Metal (MPS) GPU. If you want local embeddings and fast inference, install natively instead: pipx install 'cocoindex-code[full]'. The :latest (slim) variant is unaffected — LiteLLM runs the model on the provider's side, so Docker vs. native makes no difference.

Quick start — docker compose up -d

Bring it up in one line — no clone needed (bash / zsh):

# macOS / Windows
docker compose -f <(curl -L https://raw.githubusercontent.com/cocoindex-io/cocoindex-code/ref