prooftrail

ProofTrail for AI Agents

ProofTrail matters to AI-agent builders for one practical reason:

it gives an agent stack a governed browser execution layer with retained evidence, guided recovery, and MCP, instead of only a one-shot browser bot.

This page is intentionally narrow. It does not claim that ProofTrail is a generic AI agent platform. It explains why an AI-agent team would still care about this repo.

It is also the right page when your team is evaluating ProofTrail as the browser-evidence layer behind Codex, Claude Code, OpenHands, OpenCode, OpenClaw, or another AI-agent shell.

Use this page when

You are trying to answer one of these questions:

This page is also the right landing page if your search sounded like:

If that is your exact search intent, continue with ProofTrail for Coding Agents and Agent Ecosystems after this page.

If you are still asking “can this repo produce one truthful baseline run at all?”, start with the canonical evaluation path first:

  1. Human-first 10 minute guide
  2. Run evidence example
  3. Evidence, Recovery, and Review Workspace

What an AI agent actually gets here

AI agents do not get a magic autonomy shell from ProofTrail.

They get four concrete things:

  1. One canonical run lane
    • start from just run
    • keep one public baseline instead of many unofficial entrypoints
  2. Retained evidence bundles
    • inspect the run later through the manifest and linked proof reports
  3. Recovery and review surfaces
    • explain, share, compare, and review the run before widening handoff
  4. A governed MCP road
    • let an external AI client consume tools without turning ProofTrail into a hidden second backend

If you are using coding-agent or agent-stack workflows

ProofTrail does not claim a dedicated first-class Codex integration or a Claude-only runtime, and it does not claim first-party OpenHands, OpenCode, or OpenClaw adapters either.

The truthful fit is narrower and still useful:

That means the most honest search-intent mapping is:

If you searched for… The truthful fit here
browser automation for Codex ProofTrail can be the browser-evidence layer behind a Codex-style agent loop
browser automation for Claude Code ProofTrail can be the browser execution and evidence layer behind Claude Code style workflows
browser automation for OpenHands ProofTrail can sit behind an OpenHands-style runtime as the browser-evidence and recovery subsystem
browser automation for OpenCode ProofTrail can provide a governed browser tool layer beneath an OpenCode-style shell
browser automation for OpenClaw ProofTrail can act as the browser workflow backend behind an OpenClaw-style gateway or tool router
MCP server for coding agents ProofTrail exposes a governed MCP surface on top of the same backend and artifacts
browser automation API for AI agents ProofTrail exposes a contract-first API for direct integration

This is a strong fit for Codex-, Claude Code-, OpenHands-, OpenCode-, and OpenClaw-style agent stacks without pretending that ProofTrail is the whole coding agent.

That combination matters because most agent systems eventually hit the same moment:

the run already happened, and now you need to inspect it, compare it, recover it, or hand it to a human without losing trust.

Search phrases builders actually use

If your team is searching for this category, the truthful intent usually looks more like:

This repo is trying to win those searches honestly by being strong at browser evidence, recovery, API, and MCP, not by pretending to be a generic autonomy product.

If you arrived here from a named agent ecosystem

Many builders land here with a very direct search intent:

The truthful answer is:

That means the practical fit is:

If you need a vendor-specific plugin or a generic AI copilot shell, this repo is not pretending to be that.

The category fit in one paragraph

ProofTrail is not trying to beat generic browser agents at open-ended roaming.

It is trying to win on:

If your need is “let the browser improvise freely,” this repo is not pretending to be that.

If your need is “let the agent use browser automation without losing the evidence and recovery story,” this repo is highly relevant.

Three useful roads for AI-agent teams

1. Canonical run first

Use this road when the AI system still needs one truthful baseline.

The rule is simple:

  1. run the mainline
  2. inspect the retained evidence
  3. only then decide whether MCP or AI reconstruction should enter the picture

Start here:

2. MCP for governed tool use

Use this road when an external AI client should consume tools instead of raw REST calls.

The role split stays clean:

Continue here:

3. AI reconstruction after artifacts exist

Use this road when artifacts already exist and an agent or operator needs help rebuilding or refining a flow.

The important boundary is unchanged:

Continue here:

Where this fits in a real agent stack

ProofTrail fits best as the browser execution and evidence layer inside a larger system.

Think of the stack like this:

Layer What stays outside ProofTrail What ProofTrail contributes
Agent shell planning, tool choice, multi-step reasoning governed browser-facing capability
Browser execution raw browser scripts or helper-only replay canonical run lane and retained evidence
Recovery and judgment ad-hoc human debugging explain, share, compare, and review packet surfaces
Integration direct tool consumption by external AI clients MCP surface over the same trusted backend and evidence model

This is why ProofTrail can matter to AI agents without claiming to be the whole agent platform.

Builder path after the audience fit is clear

Once the category fit makes sense, move from positioning to contract:

  1. API Builder Quickstart
  2. Universal API Reference
  3. node --import tsx contracts/scripts/generate-client.ts --verify

That path keeps the story honest:

Suggested reading order

Use this order if you want the shortest truthful route from category fit to repo reality:

  1. README.md
  2. Run evidence example
  3. Evidence, Recovery, and Review Workspace
  4. MCP for Browser Automation
  5. AI Reconstruction Side Road
  6. ProofTrail vs Generic Browser Agents

Honest boundary

This page does not claim:

It only claims what the repo already supports: