pi-mgrep

pi-mgrep

Ripgrep + mgrep dual-routing. Both auto-install. Zero config.

Unified search extension for Pi Coding Agent.

Quick Start

npm install -g pi-mgrep
# Restart pi — ripgrep and mgrep auto-install on first use.

What it gives you

Three LLM-callable tools.

search — Dual-Engine Router

Auto-routes based on query type:

Query	Engine	Speed
registerTool (camelCase)	ripgrep	0.02s
calculate_total (snake_case)	ripgrep	0.02s
app.tsx (path/ext)	ripgrep	0.02s
{ status: 200 } (syntax)	ripgrep	0.02s
error handling logic (NL)	mgrep	3-8s
how to handle errors (NL)	mgrep	3-8s

If ripgrep is unavailable, mgrep handles everything silently.

search({ query: "registerTool" })
// → ripgrep 0.02s, file + line matches

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

web_search — Internet Search

mgrep web with automatic DuckDuckGo fallback.

web_search({ query: "React 19 new features" })
// → 5 ranked URLs

web_search({ query: "React 19 new features", count: 10 })
// → 10 ranked URLs (max customizable)

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

web_fetch — Page Reader

Fetch URL, strip HTML, return plain text.

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

Customizing Results Count

Default is 5 results per search.

Per-query (via prompt)

Ask the Agent in natural language — no code changes needed:

"Search the web for React 19 new features, return 10 results"
"请搜索 React 19 新特性，返回 10 条结果"

The Agent will automatically pass count: 10 to web_search. Works in English and Chinese.

Change the default permanently

Ask the Agent directly:

"请将 web_search 的默认返回数量从 5 改为 10"
"Change the default result count of web_search from 5 to 10"

The Agent will edit .pi/extensions/mgrep.ts for you — change both occurrences of the number, then restart pi.

Interactive Commands

Command	Purpose
/search <query> [path]	Local search (auto routes rg/mgrep)
/web <query>	Web search
/fetch <url>	Fetch URL content

How it works

┌─ resolveRg() ──────────────────────────────────────────────┐
│  which rg → known paths → brew install ripgrep              │
├─ resolveMgrep() ───────────────────────────────────────────┤
│  which mgrep → known paths → npm install -g @mixedbread/mgrep │
├────────────────────────────────────────────────────────────┤
│  Tool: search                                               │
│    code-like? → rg (if available, 0.02s)                    │
│             els → mgrep (3-8s, semantic)                    │
│             answer=true? → -a (AI summary)                  │
├────────────────────────────────────────────────────────────┤
│  Tool: web_search                                           │
│    mgrep -w /tmp/mgrep-empty                                │
│    fail? → DuckDuckGo HTML (zero-dep fallback)              │
├────────────────────────────────────────────────────────────┤
│  Tool: web_fetch                                            │
│    Node.js http.get → strip HTML → plain text               │
└────────────────────────────────────────────────────────────┘

Requirements

• Pi Coding Agent ≥ 0.73.0
• Everything else auto-installs (ripgrep + mgrep)

Authentication

mgrep needs auth to Mixedbread's API.

Device login (7-day expiry)

mgrep login

API key (recommended, permanent)

• Mixedbread Platform → sign up
• Create API key → export:

# ~/.zshrc
export MXBAI_API_KEY="mxb_your_key_here"

Design choices

• Dual-engine — ripgrep 0.02s for code, mgrep 3-8s for natural language
• Auto-install — both engines install via system package managers
• Graceful degradation — rg missing → mgrep for all. mgrep missing → DDG for web
• Output limits — 6000 char cap on every tool

Test Report

Category	Scenario	Engine	Time
Code symbol registerTool	Local	ripgrep	0.007s
NL web search fallback logic	Local	mgrep	12.9s
Chinese NL 错误处理策略	Local	mgrep	4.5s
NL + answer summary	Local	mgrep	13.7s
Default count: 5	Web	mgrep	5.1s
answer: true + 9 citations	Web	mgrep	9.9s
count: 3 / count: 7 truncation	Web	filterWeb	—
HTML fetch & strip	Fetch	Node.js	<1s
DuckDuckGo fallback	Degrade	DDG	1.3s
6000-char truncation	Edge	—	—

15/15 passed. Full details: docs/test-report.md.

License

MIT