ZeroToken: Record Once, Automate Forever - An MCP Server That Eliminates LLM Token Waste in Browser Automation

Forum topic · ✨步子哥 · 2026-03-14

Summary

ZeroToken is an open-source MCP (Model Context Protocol) server that addresses a key inefficiency in AI agent browser automation: when an AI agent controls a browser, every action requires LLM reasoning, so repeatedly executing the same task wastes massive token costs on re-deriving identical workflows. ZeroToken uses a three-layer architecture: (1) a recording phase where LLM reasoning plans the task, (2) trajectory generation that captures the workflow, and (3) script replay via a script engine that runs subsequent executions without any LLM calls. Core features include fuzzy point marking to flag steps that require human or AI judgment, deterministic replay after the first inference, adaptive element fingerprinting to survive website layout changes, and a unified structured error contract with retryable flags for AI decision-making. Compared to the traditional approach where every execution requires LLM inference, ZeroToken replays recorded scripts deterministically, significantly reducing token costs for stable, repetitive tasks. The project is available on GitHub at github.com/AMOS144/zerotoken.

ZeroToken: Record Once. Automate Forever.

ZeroToken is an AI agent browser automation MCP Server designed to eliminate redundant LLM inference in repetitive browser tasks.

The Core Problem

When an AI agent controls a browser, every operation requires one LLM inference. When the same task is executed repeatedly, the system re-thinks the identical workflow every day — causing enormous token waste.

Three-Layer Architecture

1. Recording Phase — LLM reasoning plans the task once 2. Trajectory Generation — the workflow is captured as a trajectory 3. Script Replay — a script engine replays the task with no LLM involved

Core Features

| Feature | Description | Example | |---|---|---| | Fuzzy Point Marking | Explicitly mark steps that need human or AI judgment | fuzzy_point: { requires_judgment: true } | | Deterministic Replay | After the first inference, run the script directly | run_script(task_id) // no LLM needed | | Adaptive Element Location | Save element fingerprints to cope with website changes | adaptive: true // intelligent matching | | Unified Error Contract | Structured errors for easier AI decision-making | { retryable: true, code: "..." } |

Efficiency Comparison

Traditional approach:

Every execution requires LLM inference
Repeatedly re-reasons the same workflow
High token costs

ZeroToken:

Replays directly after the first inference
Stably executes deterministic tasks
Dramatically lower token costs