Tabstack — Web & PDF Tools for AI Agents

Tabstack is a web execution API for reading, extracting, transforming, and interacting with web pages and PDF documents. It handles JavaScript-rendered sites, structured data extraction, AI-powered content transformation, and multi-step browser automation.

Setup (first use only)

Install dependencies from the skill's directory:

cd <skill-dir> && npm install

Where <skill-dir> is the directory containing this SKILL.md file.

Operations

All operations are run via the exec tool. First cd into the skill directory, then run the command with a relative path:

<skill-dir>/scripts/run.sh <command> <args>

Execution strategy: Always run tabstack commands in the foreground — call exec and wait for completion. Background execution requires manual polling and is unreliable.

JSON arguments: Any JSON argument (schema, --data) can be passed inline or as a file path prefixed with @ (e.g. @/tmp/schema.json). Use file paths for complex schemas to avoid shell quoting issues.

1. extract-markdown — Read a page or PDF as clean Markdown

Best for: reading articles, documentation, PDF reports. This is the cheapest operation — prefer it when you just need to read content.

<skill-dir>/scripts/run.sh extract-markdown "<url>"

Returns the page/PDF as Markdown. For web pages, includes YAML frontmatter metadata (title, author, etc.).

Optional flags:

• --metadata — return metadata as a separate JSON block
• --nocache — bypass caching and get fresh content
• --geo CC — fetch from a specific country (ISO 3166-1 alpha-2, e.g. US, GB)

2. extract-json — Pull structured data from a page or PDF

Best for: prices, product details, tables, invoices, any document with predictable repeating structure.

Without a schema (Tabstack infers structure):

<skill-dir>/scripts/run.sh extract-json "<url>"

With a JSON Schema (inline or from file):

<skill-dir>/scripts/run.sh extract-json "<url>" @/tmp/schema.json

Optional flags: --nocache, --geo CC.

See references/examples.md for common JSON schema patterns (products, articles, events, tables, contacts).

3. generate — Transform web/PDF content into a custom JSON shape

Best for: summaries, categorization, sentiment analysis, reformatting. Unlike extract-json (which pulls existing data), generate uses an LLM to create new content. May be slower due to LLM processing.

<skill-dir>/scripts/run.sh \
  generate "<url>" "<json_schema|@file>" "<instructions>"

Optional flags: --nocache, --geo CC.

Example — categorise and summarise HN posts:

<skill-dir>/scripts/run.sh \
  generate "https://news.ycombinator.com" \
  '{"type":"object","properties":{"stories":{"type":"array","items":{"type":"object","properties":{"title":{"type":"string"},"category":{"type":"string"},"summary":{"type":"string"}}}}}}' \
  "For each story, categorize as tech/business/science/other and write a one-sentence summary"

See references/examples.md for more schema and instruction examples.

4. automate — Multi-step browser task in natural language

Best for: tasks needing real browser interaction — clicking, navigating across pages, filling forms. Does NOT support PDFs or --geo.

<skill-dir>/scripts/run.sh \
  automate "<natural language task>" --url "<url>"

Optional flags:

• --url <url> — starting URL for the task. When omitted, automate uses its own built-in web search to find relevant pages — this can be cheaper and faster than research for simple factual questions.
• --max-iterations N — limit steps (default 50, range 1-100)
• --guardrails "..." — safety constraints (e.g. "browse only, don't submit forms")
• --data '{"key":"val"}'|@file — JSON context for form filling

Timeout: May take 30-120 seconds. Use at least 420s exec timeout.

Example — fill a contact form with guardrails:

<skill-dir>/scripts/run.sh \
  automate "Fill out the contact form with my information" \
  --url "https://example.com/contact" \
  --data '{"name":"Alex","email":"alex@example.com","message":"Hello"}' \
  --guardrails "Only fill and submit the contact form, do not navigate away"

Example — simple search (no URL, uses built-in web search):

<skill-dir>/scripts/run.sh \
  automate "Find the current price of a MacBook Air M4"

5. research — AI-powered deep web research

Searches the web, analyzes multiple sources, and synthesizes a comprehensive answer with citations. Unlike the other operations, research doesn't need a URL — you give it a question and it finds the answers.

For simple factual lookups, automate without a --url may be faster and cheaper. Use research when you need depth, multiple perspectives, or cited sources.

Use cases:

• Complex questions that need multiple sources ("What are the pros and cons of Rust vs Go for CLI tools?")
• Fact-checking and verification ("Is it true that...")
• Current events and recent information
• Topic deep-dives and literature reviews
• Competitive research ("Compare X vs Y vs Z")

<skill-dir>/scripts/run.sh research "<query>"

Optional flags:

• --mode fast|balanced — fast for quick single-source answers, balanced (default) for deeper multi-source research with more iterations
• --geo CC — research from a specific country's perspective

Timeout: May take 60-120 seconds. Use at least 420s exec timeout.

Example — quick factual lookup:

<skill-dir>/scripts/run.sh research "What is the current LTS version of Node.js?" --mode fast

Example — deep research:

<skill-dir>/scripts/run.sh research "Compare WebSocket vs SSE vs long polling for real-time web applications"

Reference: Examples & Recipes

Read references/examples.md when you need to:

• Build a JSON schema for extract-json — patterns for products, articles, events, tables, contacts, invoices
• Write effective instructions for generate — recipes for summarization, sentiment analysis, competitive analysis, content digests
• Recover from a failed attempt — if a command doesn't produce good results, check for a better approach

Choosing the Right Operation

Prefer cheaper operations when they suffice. Use extract-markdown for simple reading. Only use automate when the task requires clicking, navigating, or form interaction.

Inform the user before triggering multiple automate calls — they are the most expensive.

Error Handling

On automate failures, retry once. If it fails again, fall back to extract-markdown for read-only tasks.

Environment Configuration

This skill requires a TABSTACK_API_KEY to function. Get one from tabstack.ai (Mozilla-backed, free tier available).

Set the key via the CLI:

openclaw config set env.TABSTACK_API_KEY "your-key-here"

The skill will exit with an error if the key is not set.

Security & Privacy

• API key: This skill requires a TABSTACK_API_KEY. All requests are sent to the Tabstack API (api.tabstack.ai) using this key for authentication. The key is read from the environment, not hardcoded.

• Data sent to Tabstack: URLs you process, JSON schemas, instructions, and any --data payloads are sent to Tabstack's servers for processing. Do not pass passwords, authentication tokens, or other secrets via --data unless you explicitly trust the Tabstack service.

• Browser automation: The automate command drives a remote browser that can click, navigate, fill forms, and submit data. Use --guardrails to constrain what the browser can do (e.g. "browse only, don't submit forms").

• Dependencies: This skill installs @tabstack/sdk and tsx from npm. A package-lock.json is provided for reproducible installs.

• No persistence: The skill does not modify agent configuration, store credentials, or run outside of its own directory.