pi-mgrep

pi-mgrep

Unified search extension for Pi Coding Agent — semantic search + exact grep + web fallback in one file.

Quick Start

# 1. Install mgrep CLI
npm install -g @mixedbread/mgrep
mgrep login   # one-time device auth

# 2. Install this extension
mkdir -p ~/.pi/agent/extensions
cp .pi/extensions/mgrep.ts ~/.pi/agent/extensions/mgrep.ts

# 3. Restart pi — done

What it gives you

Three LLM-callable tools, automatically registered:

search — Local Intelligent Router

Auto-selects the right engine based on the query pattern:

Query Pattern	Engine	Speed
registerTool (camelCase)	ripgrep	0.02s
calculate_total (snake_case)	ripgrep	0.02s
app.tsx (file extension)	ripgrep	0.02s
how to handle errors (NL)	mgrep	3-8s

// LLM calls this automatically:
search({ query: "registerTool" })
// → ripgrep 0.02s, returns file + line matches

search({ query: "error handling logic", answer: true })
// → mgrep semantic search, returns AI-generated summary

web_search — Internet Search

Semantic search the web via mgrep, with automatic DuckDuckGo fallback.

web_search({ query: "React 19 new features" })
// → 3 ranked URLs with match scores

web_search({ query: "React 19 new features", answer: true })
// → AI-generated answer with citations instead of URL list

Fallback chain: mgrep web (primary) → DuckDuckGo HTML scraping (automatic on failure)

web_fetch — Page Reader

Fetch a URL, strip HTML, return clean text.

web_fetch({ url: "https://react.dev/blog/2024/12/05/react-19" })
// → Clean text content, max 6000 chars

Interactive Commands

Command	Purpose
/search <query> [path]	Local search with auto engine selection
/web <query>	Web search with AI summary
/fetch <url>	Fetch and display URL content

Architecture

┌─ search (local) ──────────────────────────────────────────┐
│  Intelligent routing:                                      │
│    code patterns → ripgrep (0.02s, offline)                │
│    natural lang  → mgrep   (3-8s, semantic)                │
│  Fallback: mgrep ↓ → ripgrep                               │
├─ web_search (internet) ───────────────────────────────────┤
│  Primary:  mgrep web   (semantic + reranking)              │
│  Fallback: DuckDuckGo  (HTML scraping, no API key needed)  │
│  Isolation: points to /tmp/mgrep-empty (no local mixing)   │
├─ web_fetch (page reader) ─────────────────────────────────┤
│  curl + Node.js HTML→text extraction                       │
│  No external dependencies                                   │
└────────────────────────────────────────────────────────────┘

Requirements

• Pi Coding Agent ≥ 0.73.0
• mgrep CLI (npm install -g @mixedbread/mgrep)
• ripgrep (optional — local semantic search auto-upgrades to mgrep if unavailable)

Authentication

mgrep requires authentication to Mixedbread's API. Two options:

Option A: Device login (7-day expiry)

mgrep login
# opens browser → authorize → token saved to ~/.mgrep/token.json
# must re-login every 7 days

Option B: API key (recommended, permanent)

• Go to Mixedbread Platform
• Sign up / Sign in
• Create an API key from the dashboard
• Export it in your shell profile:

# ~/.zshrc (or ~/.bashrc)
export MXBAI_API_KEY="mxb_your_key_here"

API key takes priority over device login and never expires. Ideal for CI/CD and daily use.

Tokens

Adds ~400 tokens to system prompt (3 tool definitions). Negligible on 128K+ context windows.

File Structure

pi-mgrep/
├── .pi/
│   └── extensions/
│       └── mgrep.ts        ← the extension (349 lines)
├── package.json
├── README.md
└── README_CN.md

Related

• mgrep — Semantic grep CLI
• Pi Extensions — Official docs
• Mixedbread Platform — API key management

License

MIT