codeindex

A structural code intelligence engine that runs as an MCP server for AI coding agents.

It indexes your codebase with tree-sitter (40+ languages), builds a trigram full-text index, an inverted word index, and a dependency graph — then exposes them through 16 MCP tools.

 ┌─────────────┐      MCP (stdio)       ┌──────────────┐
 │  AI Agent   │ ◄─────────────────────► │   codeindex  │
 │ (Claude,    │   16 tools, JSON-RPC    │  (Zig binary) │
 │  Cursor…)   │                         │              │
 └─────────────┘                         └──────┬───────┘
                                                │
                                   tree-sitter parse (40+ langs)
                                         trigram + word index
                                         dependency graph
                                         snapshot persistence

Why

AI coding agents spend tokens reading entire files. codeindex answers structural questions — symbol outlines, definitions, callers, blast radius, dependency chains — in a few hundred tokens instead of thousands.

One plan_change call returns: where a symbol is defined, every call site, the file's architectural role (god module / stable core / island / driver), hardcoded literals to check, and the full transitive blast radius if the file changes.

Related MCP server: OpenCodeHub MCP Server

Quickstart

The shortest path, if you have Node 18+. Nothing else to install, no key, no config — the package is a 4 KB wrapper that fetches the binary for your platform and verifies it against the published checksums:

claude mcp add codeindex -- npx -y @munhq/codeindex

No Node, or you want the skill and the hook as well:

# Prebuilt binary + skill + MCP registration, in one command
curl -fsSL https://raw.githubusercontent.com/munhq/codeindex/main/install.sh | sh

# Or build from source:
cd zig && ./fetch-vendor.sh && zig build -Doptimize=ReleaseFast

Docker, for hosts that install MCP servers as images. The workspace is bind-mounted read-only; codeindex never writes to it:

docker run -i --rm -v "$PWD:/workspace:ro" munhq/codeindex

Register with your AI agent:

# Claude Code — the plugin is the one-step path. It ships the skill, both
# routing hooks and the MCP server together, and its launcher finds or fetches
# the binary.
claude plugin marketplace add munhq/codeindex
claude plugin install codeindex@codeindex

# Without the plugin (or for a different MCP client), register the binary
# directly. Do not do both: two registrations mean two servers, two copies of
# every tool schema, and two writers on one snapshot. install.sh detects the
# plugin and skips this step when it is present.
claude mcp add -s user codeindex -- ~/.local/bin/codeindex --mcp

# Cursor / Claude Desktop / other MCP clients: add to your config
{
  "mcpServers": {
    "codeindex": {
      "command": "npx",
      "args": ["-y", "@munhq/codeindex"]
    }
  }
}

Every listing points at the same server: npm @munhq/codeindex, the official MCP registry as io.github.munhq/codeindex, and Smithery as munhq/codeindex.

The next time your agent starts, codeindex indexes your project in the background and serves structural queries.

MCP Tools

Analyses (analyze tool)

Supported Languages

40+ languages via tree-sitter: Rust, Python, TypeScript/TSX, Go, Zig, C, C++, Java, Ruby, Bash, C#, Kotlin, Lua, Scala, Elixir, R, Swift, Dart, Haskell, TOML, JSON, YAML, HTML, CSS, SCSS, SQL, HCL, Dockerfile, Markdown, Nix, Make, and more.

Configuration

codeindex --mcp                          # Run as MCP server (stdio)
codeindex --workspace ./my-project       # Index a specific directory
codeindex --project-id my-project        # Project identifier
codeindex -v                             # Print version
codeindex -h                             # Print help

# Environment variables
CODEINDEX_WORKSPACE=/path/to/project     # Same as --workspace
CODEINDEX_PROJECT_ID=my-project          # Same as --project-id

Getting it used

A server that registers without telling an agent when to reach for it stays idle, and an idle index saves nothing however cheap its calls are. The plugin ships three things for that, in the order they act:

1. A SessionStart brief. About 240 tokens, once per session, in a repository that holds source files: codeindex is live, and here is the tool for each kind of code question. It lands before the agent has chosen a tool, which is the only moment that can change the first choice.

2. A PreToolUse hint. Fires on the 1st, 8th and 25th code question of a session, when a scan is about to answer something the index answers better, and names the tool for that exact question. It matches Bash as well as Read/Grep/Glob, because a permission mode that routes file work through the shell is where most scans actually happen.

3. The skill, which the model loads when it decides the task calls for it.

Advice loses to habit, so the narrow case where the index is strictly better is refused rather than argued with: an identifier search that is the whole command line, in a project that has been indexed before. It never refuses a file read — the bytes are required before an Edit — never refuses a line that also builds or tests, and stops refusing after three times in a session, so it can never be the reason a session cannot proceed.

That is on by default, and it is one switch to turn off — no file to edit by hand. It is a plugin option, so /plugin shows it as "Refuse a scan the index answers better" and Claude Code passes your answer to the hook. Outside the plugin, or for a single session, CODEINDEX_ENFORCE=0 in the environment overrides the option and CODEINDEX_ENFORCE=1 restores it.

Measure the effect rather than assuming it: python3 plugin/measure_routing.py counts, per session, how many code questions went to a scan and how many went to the index, across every Claude Code home on the machine.

Which tree gets indexed

In --mcp mode the working directory belongs to the client, not to you, and it is regularly not your project: Claude Code launches a plugin's MCP server from the plugin's own directory and a user-scope server from its config directory. So the workspace is resolved in this order, and the first usable answer wins:

1. --workspace or CODEINDEX_WORKSPACE. An explicit answer is never overridden.

2. CLAUDE_PROJECT_DIR, which Claude Code sets in every MCP server it spawns.

3. The launch directory, walking up for a .git, package.json, Cargo.toml, go.mod, build.zig or pyproject.toml marker.

4. The client's MCP roots. The server asks for them at the handshake whenever the root above was a guess, and adopts one when the guess was refused or lands outside every root the client reports. notifications/roots/list_changed is honoured too, so a directory added with --add-dir can still rescue a session.

status reports the effective root as workspace, and whether a file watcher is running as watcher. Check them when a result looks empty: an index of the wrong tree answers every question just as confidently as the right one.

It refuses to index your entire home directory, the filesystem root, or a folder that merely holds several independent repositories. A refused workspace answers every query with the reason and the one call that fixes it — index_workspace with the project path — rather than with "no results".

Architecture

• Parser: tree-sitter with 40+ grammars, compiled into a single binary

• Index: trigram index for fuzzy text search + inverted word index for exact identifier lookup

• Dependency graph: file-level import resolution with forward and reverse edges

• Version store: tracks file changes with sequence numbers for incremental updates

• Live watcher: re-indexes on file create/modify/delete (background thread in MCP mode). inotify on Linux; a polling walk on macOS and Windows, which compares mtime and size every couple of seconds. status reports which backend is live as watcher_backend.

• Snapshot: persists the full index to .codeindex.json, so a restart loads the snapshot instead of re-indexing

• MCP server: JSON-RPC over stdio, implements the MCP 2024-11-05 protocol

Platform support

Every row is built by CI and its tests are run on that platform, except where noted. status reports the live watcher backend so it is never a guess.

On Windows, install.sh and the plugin's launcher are shell scripts, so they need Git Bash, MSYS2 or Cygwin — they detect it and resolve the right .exe asset. The plugin registers its server through that launcher, so a native Windows Claude Code without a shell should register the binary directly:

claude mcp add -s user codeindex -- C:\path\to\codeindex.exe --mcp

Nothing here is signed or notarized. On macOS a binary fetched with curl runs without a Gatekeeper prompt; one downloaded through a browser is quarantined, and xattr -d com.apple.quarantine codeindex clears it.

Building from source

Requires Zig 0.16.0.

cd zig
./fetch-vendor.sh    # Clone tree-sitter + 40 grammar repos
zig build -Doptimize=ReleaseFast
# Binary: zig/zig-out/bin/codeindex

Run tests:

cd zig && zig build test-bin && ./zig-out/bin/test

zig build test routes results through the build runner's IPC protocol on stdout, which the linked tree-sitter C sources corrupt via their debug printf paths. Building the test binary and running it directly is the same tests without that protocol in the way.

How it compares

Pairs with chat-recall

codeindex answers questions about the code in front of you. chat-recall answers questions about the work you already did — it indexes your Claude Code, Gemini CLI, Codex, OpenCode and Antigravity sessions into one searchable history and exposes that over MCP too.

Together they cover both halves of what an agent forgets: codeindex stops it re-reading files it could have outlined, and chat-recall stops it redoing work it already finished. chat-recall detects a codeindex binary on your PATH and registers four extra code-intelligence tools when it finds one — neither requires the other.

License

MIT. See LICENSE.