Agent framework integrations

Tabstack ships maintained packages that expose its capabilities as native tools for common agent frameworks. Instead of hand-rolling a tool around the SDK, you install one package and drop a ready-made tool set into your agent.

Every package wraps the official Tabstack SDK and exposes the same five tools, with the same names, descriptions, and inputs, so behavior is consistent across languages and frameworks.

Choose your package

Framework	Language	Package	Guide
LangChain	Python	langchain-tabstack	LangChain (Python)
LangChain.js	TypeScript	@tabstack/langchain	LangChain.js
Vercel AI SDK	TypeScript	@tabstack/ai	Vercel AI SDK
Mastra	TypeScript	@tabstack/mastra	Mastra
LlamaIndex.TS	TypeScript	@tabstack/llamaindex	LlamaIndex.TS
OpenAI Agents SDK	TypeScript	@tabstack/openai-agents	OpenAI Agents SDK
Claude Agent SDK	TypeScript	@tabstack/claude-agent	Claude Agent SDK
eve	TypeScript	@tabstack/eve	eve

If you are not using one of these frameworks, call the TypeScript or Python SDK directly.

The tools

Tool	Backed by	Use it for
extract_structured_data	extract.json	Pull specific fields from a URL into a JSON shape you define.
extract_page_content	extract.markdown	Fetch a page as clean markdown.
research_question	agent.research	Synthesize a cited answer across multiple pages.
generate_structured_data	generate.json	Fetch a page, then AI-transform it into derived or reshaped JSON.
automate_browser_task	agent.automate	Run a multi-step, natural-language browser task.

Shared concepts

These apply to every package.

Optional inputs

Beyond each tool’s required inputs, the fetch-based tools accept optional controls. They are sent to Tabstack only when provided, so omitting them keeps the defaults.

• extract_structured_data, extract_page_content, generate_structured_data:
  • effort: min, standard, or max. Use max for JS-heavy pages (full server-side browser rendering). See Effort levels.
  • nocache: bypass the cache and fetch fresh.
  • country: ISO 3166-1 alpha-2 code for geotargeting.
• research_question: mode (fast or balanced), nocache.
• automate_browser_task: guardrails (constraints on what the agent may do), data (context for form filling), country, max_iterations, max_validation_attempts.

Browser automation runs non-interactively

automate_browser_task consumes the automate stream without the human-in-the-loop form-data flow, so it never blocks waiting for input. It returns the final answer plus the data it extracted and the pages it visited. For interactive automation, use the SDK’s interactive mode directly.

Output shapes

The TypeScript packages return native objects (camelCase keys). The Python package returns JSON strings (snake_case keys), the conventional LangChain-Python tool output. The data is the same; only the return type and key casing differ, by language convention.

@tabstack/claude-agent is the one structural exception: it registers the tools as an in-process MCP server rather than handing you tool objects, so results reach the model as MCP text content (markdown for extract_page_content, JSON for the rest).

Why use Tabstack here

• Schema-enforced output. You define the shape, you get that shape back. No prompt-engineering the JSON out of raw text.
• No browser to run. JS-heavy pages render server-side. No Playwright, no binaries.
• Version-independent. It is an SDK call, not a loader coupled to your framework version. See Tabstack vs. LangChain browser tools.