# haiku.rag MCP server

> Search and question your own documents with hybrid retrieval, reranking and answers cited to page numbers and headings.

- Listing: https://mcp.tc/i/haiku-rag
- Connect: this is a local (stdio) server; install it on your machine (see Install). The listing link is a page, not an MCP endpoint.
- Type: local (stdio)
- Auth: no API key needed
- Category: [Docs & Knowledge](https://mcp.tc/c/docs-knowledge)
- Vendor: ggozad
- Homepage: <https://ggozad.github.io/haiku.rag/>
- Docs: <https://ggozad.github.io/haiku.rag/>
- Repository: <https://github.com/ggozad/haiku.rag>
- Package: pypi `haiku-rag`

## About

haiku.rag is a local-first agentic RAG system built on embedded LanceDB, Pydantic AI and Docling. It indexes PDFs, office documents, HTML, Markdown and images, then lets an assistant search them, read documents and run sandboxed Python over them. Answers cite page numbers and section headings.

It runs locally over stdio with the command haiku-rag mcp --stdio, installed from the PyPI package haiku-rag (run with uvx). It needs Python 3.12 or newer. No server or API key is needed by default, since it uses open models through Ollama. Other providers supported by Pydantic AI can be used instead.

## What it can do

- Hybrid vector and full-text search over indexed documents
- Optional reranking and section-aware context expansion
- Read indexed documents with page and heading provenance
- Run sandboxed Python over documents for counting and aggregation
- Ingest PDFs, office files, HTML, Markdown and images
- Multimodal image search with a supported embedder

## Example prompts

- "Search my indexed papers for the attention mechanism and cite page numbers."
- "What datasets were used for evaluation in paper.pdf?"
- "How many documents in my knowledge base mention transformers?"
- "Read the section on methodology from the indexed spec document."

## Install

### Claude Code

1. Run this in a terminal, in your project folder:

```bash
claude mcp add --transport stdio haiku-rag -- uvx haiku-rag mcp --stdio
```

2. Start Claude Code and type `/mcp`. **haiku-rag** should show as connected.

Add `--scope user` to make it available in every project.

### Claude Desktop

1. Open **Settings → Developer → Edit Config**. It opens `claude_desktop_config.json`. Add:

`claude_desktop_config.json`:

```json
{
  "mcpServers": {
    "haiku-rag": {
      "command": "uvx",
      "args": [
        "haiku-rag",
        "mcp",
        "--stdio"
      ]
    }
  }
}
```

2. Save the file and restart Claude Desktop.

Needs uv (Python) on your computer. The file lives in `~/Library/Application Support/Claude/` on macOS and `%APPDATA%\Claude\` on Windows.

### Cursor

[Add to Cursor](<https://cursor.com/install-mcp?name=haiku-rag&config=eyJjb21tYW5kIjoidXZ4IiwiYXJncyI6WyJoYWlrdS1yYWciLCJtY3AiLCItLXN0ZGlvIl19>) (opens Cursor)

Or add it by hand to `~/.cursor/mcp.json` (all projects) or `.cursor/mcp.json` (this project):

`mcp.json`:

```json
{
  "mcpServers": {
    "haiku-rag": {
      "command": "uvx",
      "args": [
        "haiku-rag",
        "mcp",
        "--stdio"
      ]
    }
  }
}
```

Needs uv (Python) on your computer.

### VS Code

[Install in VS Code](<https://vscode.dev/redirect/mcp/install?name=haiku-rag&config=%7B%22type%22%3A%22stdio%22%2C%22command%22%3A%22uvx%22%2C%22args%22%3A%5B%22haiku-rag%22%2C%22mcp%22%2C%22--stdio%22%5D%7D>) (opens VS Code)

Or from a terminal:

```bash
code --add-mcp '{"name":"haiku-rag","type":"stdio","command":"uvx","args":["haiku-rag","mcp","--stdio"]}'
```

Or commit it to the repo in `.vscode/mcp.json`:

`.vscode/mcp.json`:

```json
{
  "servers": {
    "haiku-rag": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "haiku-rag",
        "mcp",
        "--stdio"
      ]
    }
  }
}
```

Needs uv (Python) on your computer.

### Devin Desktop

1. Add it to `~/.config/devin/mcp_config.json` (macOS and Linux) or `%APPDATA%\devin\mcp_config.json` (Windows):

`mcp_config.json`:

```json
{
  "mcpServers": {
    "haiku-rag": {
      "command": "uvx",
      "args": [
        "haiku-rag",
        "mcp",
        "--stdio"
      ]
    }
  }
}
```

2. Refresh the MCP server list in Cascade.

Devin Desktop is the new name for Windsurf.

### Codex

```bash
codex mcp add haiku-rag -- uvx haiku-rag mcp --stdio
```

Or edit `~/.codex/config.toml` directly:

`config.toml`:

```toml
[mcp_servers.haiku-rag]
command = "uvx"
args = ["haiku-rag", "mcp", "--stdio"]
```

Needs uv (Python) on your computer.

### Gemini CLI

```bash
gemini mcp add haiku-rag uvx -- haiku-rag mcp --stdio
```

This adds it to the current project. Add `-s user` to use it everywhere.

### Any client

Most clients that start local servers accept this shape:

```json
{
  "mcpServers": {
    "haiku-rag": {
      "command": "uvx",
      "args": [
        "haiku-rag",
        "mcp",
        "--stdio"
      ]
    }
  }
}
```

Zed puts servers under `context_servers` in its settings, with the same `command`, `args` and `env` fields.

Needs uv (Python) on your computer.

## Recent changes

- 2026-11-17 08:29 UTC: New version 0.112.1 (was 0.112.0), read from PyPI. Source: automatic check.
- 2026-11-13 17:26 UTC: New version 0.112.0 (was 0.111.1), read from PyPI. Source: automatic check.
- 2026-11-12 14:26 UTC: New version 0.111.1 (was 0.111.0), read from PyPI. Source: automatic check.
- 2026-11-11 08:29 UTC: New version 0.111.0 (was 0.110.0), read from PyPI. Source: automatic check.
- 2026-11-08 17:32 UTC: New version 0.110.0 (was 0.109.0), read from PyPI. Source: automatic check.

More on the History tab, with descriptions before and after for the newest entries: https://mcp.tc/i/haiku-rag#history (Atom feed: https://mcp.tc/i/haiku-rag/history.xml). It records what mcp.tc's anonymous checks saw, every 3 hours where the tool list can be read without sign-in; it is not a security review.

## Details

- Server version: 0.112.2
- Last checked: 2026-10-12
- Listed: 2026-10-12
- Updated: 2026-10-26

---
Source: https://mcp.tc/i/haiku-rag (mcp.tc is an independent directory, not affiliated with this server's publisher). Corrections: https://mcp.tc/report
