Summary

ZeroToken addresses the high token cost and slow execution of repetitive browser automation tasks for AI agents. It provides a lightweight MCP that records actions once and enables permanent, fast automation, significantly reducing LLM inference needs and operational costs.

How It Works

ZeroToken functions as a persistent HTTP-based browser execution layer, maintaining browser state across calls, unlike stdio modes that reset sessions. It records detailed "trajectories" of user actions, including page states, screenshots, and identifies "fuzzy points" requiring AI judgment. These trajectories can be converted into deterministic scripts for fast, low-token replay, bypassing LLM reasoning for repetitive tasks. AI agents interact via MCP calls, receiving structured operation records and AI-friendly prompts for further analysis.

Quick Start & Requirements

Install via pip install zerotoken and playwright install chromium. For stateful operation with OpenClaw, run zerotoken-mcp-http in the background and configure openclaw.json with the HTTP URL. Stdio mode is available for IDEs like Cursor via zerotoken-mcp.

Highlighted Details

• Stateful Browser Sessions: Achieved via a persistent HTTP MCP server, crucial for multi-step automation.
• Trajectory Recording: Captures detailed OperationRecords with steps, page state, screenshots, and optional fuzzy_point markers for AI judgment.
• Script Engine: Enables deterministic, LLM-free replay of recorded tasks from SQLite storage, ideal for scheduled jobs.
• Stability Enhancements: Features SmartSelector (prioritized selectors, unstable class filtering), SmartWait (cascading wait conditions), and ErrorRecovery (retries, selector variants).
• Anti-bot Measures: Includes stealth=true option in browser_init for fingerprint spoofing and reduced detection.

Maintenance & Community

The project welcomes community contributions via Issues and PRs. A QQ group (ID: 942359087) is available for discussion and collaboration.

Licensing & Compatibility

Released under the permissive MIT License, allowing for broad compatibility with commercial and closed-source applications.

Limitations & Caveats

Stateful browser sessions strictly require the HTTP mode; stdio mode is stateless. Initial anti-bot capabilities focus on stealth and fingerprinting, with advanced verification potentially requiring further development.