Why Pi + Tabstack

Wiring web access into a coding agent usually means hand-rolled fetches, HTML parsing, and prompt gymnastics to coax structured data out of messy text. @tabstack/pi-agent replaces that with a hosted API registered as native Pi tools through a Pi extension: schema-enforced output, server-side rendering of JS-heavy pages, and one key for extraction, research, generation, and automation.

Quickstart

Install the adapter and set your key:

npm install @tabstack/pi-agent @earendil-works/pi-coding-agent typebox zod
 
export TABSTACK_API_KEY="your-key-here"

@earendil-works/pi-coding-agent (v0.82 or later), typebox (v1), and zod (v4) are peer dependencies.

A Pi extension is a .ts file with a default export that Pi calls with its extension API. Re-export this package's default extension and Pi registers all five tools:

// tabstack.ts
export { default } from "@tabstack/pi-agent";

Then load it with the pi CLI:

pi -e ./tabstack.ts

Pi also auto-discovers extensions dropped into ~/.pi/agent/extensions/*.ts (global, every project) or .pi/extensions/*.ts (project-local, checked into the repo), with no -e flag needed. The tools resolve TABSTACK_API_KEY lazily on first call, so loading the extension never requires a key.

Registering a subset

You often do not want every tool. A coding agent that should read and research but never drive a browser can register just those two, in the order given:

// tabstack.ts
import { createTabstackPiExtension } from "@tabstack/pi-agent";
 
// Read-and-research only, no browser automation.
export default createTabstackPiExtension({
  tools: ["extract_page_content", "research_question"],
});

For full control, register individual tools in a hand-written extension. Every tool is exported as a Pi ToolDefinition, and createTabstackPiTools(config) returns them keyed by name for a custom client:

import { extractPageContentTool, researchQuestionTool } from "@tabstack/pi-agent";
 
export default function (pi) {
  pi.registerTool(extractPageContentTool);
  pi.registerTool(researchQuestionTool);
  // ...and your own tools alongside them.
}

The tools

Tool name	Label	What it does
`extract_structured_data`	Extract Structured Data	Pull specific fields from a URL into a JSON shape you define.
`extract_page_content`	Extract Page Content	Fetch a page as clean markdown.
`research_question`	Research Question	Synthesized answer with cited sources across multiple pages.
`generate_structured_data`	Generate Structured Data	Fetch a page, then AI-transform it into derived or reshaped JSON.
`automate_browser_task`	Automate Browser Task	Run a multi-step, natural-language browser task.

The names stay in lockstep with the @tabstack/langchain, @tabstack/openai-agents, @tabstack/eve, and Python langchain-tabstack packages.

Good to know

Cancellation works. Pi's abort signal is threaded into the Tabstack request, so cancelling a tool call, for example a long research_question or automate_browser_task, aborts the request and stops billing rather than just the agent loop.
automate_browser_task runs non-interactively. It does not pause for human-in-the-loop form input, so it never blocks. It returns the final answer plus the data it extracted and the pages it visited.
Failed calls throw TabstackToolError, a normalized message plus an HTTP status for API errors. Pi surfaces the failure to the model as a tool result rather than aborting the session.
Requires Zod 4. The adapter uses Zod 4's native z.toJSONSchema to derive each tool's Typebox parameters.

Common use cases

Give a coding agent cited research without leaving the terminal.
Read a page as clean markdown mid-task, with no browser to install.
Register a read-only subset so the agent can research but never act.
Check the extension into .pi/extensions/ so the whole team gets the tools.

Pi Agent