A portable agent skill (the
SKILL.md format) that gives your agent web browsing, data extraction, content
transformation, web research, and browser automation via the
Tabstack API.
Works with any runtime that reads the SKILL.md format — including
Claude Code,
OpenClaw,
nanobot, and
Hermes.
The skill drives the official tabstack CLI
— a single standalone Go binary — by shelling out to it directly.
| Command | Purpose | Cost | Timeout |
|---|---|---|---|
extract markdown |
Read a page or PDF as clean Markdown | Lowest | 60s |
extract json |
Pull structured data from a page or PDF | Medium | 60s |
generate json |
AI-transform web/PDF content into JSON | Medium | 60s |
agent research |
AI-powered multi-source web research | Medium | 420s |
agent automate |
Multi-step browser automation or web search | Highest | 420s |
The extract, generate, and agent automate commands support --geo CC for
region-specific content; agent research does not. Extract and generate also
take --effort min|standard|max to trade cost against thoroughness.
The agent automate command supports --guardrails (safety constraints) and
--data (JSON context for form filling). When called without --url, it
uses built-in web search — useful for simple factual lookups. A paused
automation can be resumed or cancelled with agent input <request-id>.
JSON arguments (schemas, --data) can be passed inline or as a file path
prefixed with @ (e.g. @/tmp/schema.json).
- A
SKILL.md-compatible agent runtime (e.g. Claude Code, OpenClaw, nanobot, Hermes) - The
tabstackCLI installed and on the agent's PATH (prebuilt binary,go install, or build from source) - A Tabstack API key
Each runtime has its own skills directory — drop the skill into the one your runtime scans. Common locations:
- Claude Code:
~/.claude/skills/tabstack/ - OpenClaw:
~/.openclaw/workspace/skills/tabstack/(per-agent workbench, highest priority) or~/.openclaw/skills/tabstack/(shared, used byclawhub install)
The examples below use $SKILLS_DIR — set it to your runtime's path, e.g.
export SKILLS_DIR=~/.claude/skills.
Copy the skill files into your runtime's skills directory:
mkdir -p "$SKILLS_DIR/tabstack"
cp -r SKILL.md references/ "$SKILLS_DIR/tabstack/"Make sure the tabstack CLI is installed and on PATH (see
Requirements), then set your API key — see
the SKILL.md Setup section for the options (tabstack auth login,
TABSTACK_API_KEY, or a runtime-specific mechanism). Reload/restart your
runtime so it picks up the new skill.
mkdir -p "$SKILLS_DIR/tabstack"
unzip tabstack.skill -d "$SKILLS_DIR/tabstack/"Validate the skill structure:
make validatePackage for distribution:
make packageThis creates tabstack.skill (a zip archive containing only the skill files)
that can be shared with other users.
Install locally (defaults to the OpenClaw workspace; override SKILLS_DIR for
another runtime):
make install # ~/.openclaw/workspace/skills/tabstack
make install SKILLS_DIR=~/.claude/skills # Claude Code- Your agent runtime discovers
SKILL.mdin its skills directory - The skill's
descriptiontells the agent when to use it - When triggered, the agent reads SKILL.md for instructions
- The agent calls
tabstack <command>directly via theexectool - The
tabstackCLI talks to the Tabstack API and returns results to stdout
SKILL.md — Skill definition and agent instructions
references/
examples.md — JSON schema patterns and generate recipes
The agent invokes the tabstack CLI directly; the skill ships no wrapper
code or dependencies of its own.
Makefile — validate, package, install targets
validate_skill.py — SKILL.md frontmatter validator
README.md — This file
LICENSE — MIT