pi-parallel

A pi extension that gives agents web access through the Parallel Search and Extract APIs.

Tool	Use case
web_search	Discover sources and find current information
web_fetch	Fetch clean markdown from known public webpages

The tools call Parallel's V1 REST API directly. No CLI process, MCP server, or SDK is involved.

Install

pi install git:github.com/HazAT/pi-parallel

Or clone it manually:

git clone https://github.com/HazAT/pi-parallel ~/.pi/agent/extensions/pi-parallel

Authentication

The extension shares authentication with parallel-cli at:

~/.config/parallel-web-tools/auth.json

If you already ran parallel-cli login, the extension uses the API key for the CLI's currently selected organization automatically.

Otherwise, start pi and run:

/parallel-setup

Paste an API key from platform.parallel.ai. The key is hidden while entering it and saved to the shared auth file with 0600 permissions. Run the command again to replace the key.

When a tool is called without authentication, it returns a clear error instructing the agent to ask you to run /parallel-setup.

How it works

• web_search sends synchronous POST /v1/search requests in low-latency Turbo mode with ranked, LLM-optimized excerpts.
• web_fetch sends synchronous POST /v1/extract requests and reports both successful pages and per-URL failures.
• Requests use the configured key through the x-api-key header and support cancellation.
• Tool output is capped at 50KB or 2,000 lines.
• Collapsed tool rows stay compact; expanded rows show page content, warnings, and fetch failures.

The agent chooses between the tools using their built-in descriptions and prompt guidelines. Typical investigation flow is web_search to find sources, followed by web_fetch for the most relevant pages.

License

MIT