zerotoken_skill

Overview

This skill integrates ZeroToken MCP with OpenClaw to enable low-token browser automation, trajectory recording, and replayable scripts. It is designed for recurring or scheduled browser tasks where minimizing LLM token use is important. It guides setup, record→script conversion, and robust scheduled execution with job_id bindings.

How this skill works

The skill uses ZeroToken MCP tools (browser_*, trajectory_*, script_*, session_*) to record user browser operations, store trajectories in the MCP SQLite backend, and convert trajectories into compact scripts. For scheduled runs OpenClaw triggers scripts by job_id via run_script_by_job_id, keeping execution token-light because playback is handled by the MCP ScriptEngine rather than an LLM. It also supports stealth mode and DFU/paused-resume flow for human-in-the-loop steps.

When to use it

• You need OpenClaw + ZeroToken for browser automation that will be repeated or scheduled.
• You already have a recorded browser trajectory and want a low-token reusable script.
• You must run unattended periodic jobs with auditability and session logs.
• You need anti-detection (stealth) for sites likely to block automation.
• Do not use for one-off ad-hoc tasks with no reuse requirement.

Best practices

• Run ZeroToken MCP in streamable HTTP mode (zerotoken-mcp-http) and configure openclaw.json to use the MCP URL so browser state persists.
• Always call browser_init before other browser_* tools and browser_close when done.
• Only convert a trajectory to a script when the task is intended to be reused or scheduled to avoid extra token work.
• Bind scheduled tasks using job_id via script_binding_set and prefer run_script_by_job_id for one-step execution.
• Include default_vars or provide vars at runtime for scripts with {{placeholders}}; otherwise execution may fail or behave unexpectedly.

Example use cases

• Daily login and report pull: record once, trajectory_to_script, bind job_id, then schedule run_script_by_job_id.
• Hourly scraping of a table: record navigation + extraction, save as script with stealth if needed, run via cron-bound job_id.
• Periodic form submission (e.g., status update): generate script, include parameterized vars for each run.
• Audit and replay failed runs: use session_list/session_get to inspect runs and re-run by task_id after fixes.