Multi-agent web research for pi. Uses a high-fidelity stealth browser system to search the web, scrape pages, check security databases, and shows everything in real-time.
High-Fidelity Stealth Research — Uses camoufox (stealth Firefox) to bypass bot detection. No search API keys or external infrastructure required.
Multi-agent Orchestration — AI coordinator breaks your question into parallel research tracks. Each track runs in a separate researcher session. An AI evaluator combines results or launches deeper rounds.
Safe by design — Researcher agents cannot write files, edit files, or run shell commands. Web tools are isolated and rate-limited to keep agents focused.
Minimal setup — Just install. A system prompt guides pi on when and how to use the tool. No advanced prompting required. Works alongside other tools without conflicts.
- Web Search — Multi-threaded, parallel search bursts using DuckDuckGo Lite.
- URL Scraping — Configurable batch scraping protocol (1-16 batches or unlimited) with PDF support and global deduplication.
- Security Databases — NVD, CISA KEV, GitHub Advisories, and OSV.
- Stack Exchange — Full network search and filtering.
- Real-time TUI — Live progress tracking with token and cost monitoring.
- Local Context — Integrated
ripgrepfor searching local codebases.
- Node.js >= 22.13.0
- pi CLI installed and configured
- Internet access
- LLM in pi with 100k+ context window
pi install npm:@lincoln504/pi-researchThis installs dependencies and the stealth browser engine. Takes a few minutes on first install.
Local install (from repo):
pi install .Just talk to pi — the research tool registers automatically, no special slash command needed.
research the latest developments in WebAssembly
deep research AI inference hardware landscape
deep research CVE-2024-3094 at depth 3
Say research for a quick lookup. Say deep research for thorough investigation — pi selects depth 1–3 based on your query's scope and complexity, or pin it with at depth N.
A /research <query> slash command is also available as a shortcut — it runs quick mode (depth 0) directly.
Depth levels:
| Depth | Mode | Researchers | Rounds |
|---|---|---|---|
| 0 | Quick | 1 | 1 |
| 1 | Normal | 2 | 2 |
| 2 | Deep | 3 | 3 |
| 3 | Ultra | 5 | 5 |
| Variable | Default | Description |
|---|---|---|
PI_RESEARCH_MAX_CONCURRENT |
3 |
Max concurrent researchers (1-10). |
PI_RESEARCH_BROWSER_HEADLESS |
true |
Run research browser in headless mode. |
PI_RESEARCH_EXPORT_DIR |
cwd |
Directory for final research reports. |
PI_RESEARCH_VERBOSE |
— | Set to 1 for diagnostic logs. |
PROXY_URL |
— | Proxy for outgoing requests (e.g., socks5://127.0.0.1:9050). |
STACKEXCHANGE_API_KEY |
— | Stack Exchange API key (increases limit). |
PI_RESEARCH_RESEARCHER_TIMEOUT_MS |
240000 |
Per-researcher timeout (default 4m). |
Commands
npm run lint/npm run lint:fix— Code qualitynpm run type-check— TypeScript verificationnpm run test:unit— Unit testsnpm run test:integration— Integration testsnpm run test:coverage— Coverage report
See docs/ARCHITECTURE.md for detailed design information.