Why AI agents need visual documentation — not just automation
WebMCP is shipping in Chrome 146. Your AI agents can now automate browser tasks natively, without external tools.
This is great. But it solves one problem and creates another.
The native automation problem
Your agent can navigate, fill forms, click buttons — all natively. But when something goes wrong, or when compliance asks "what exactly did this agent do on Jan 15 at 3 PM?", you have nothing.
No screenshot. No video. No audit trail. Just logs that say "agent.click() succeeded."
In production, that's not enough.
The three things you actually need
1. Visual proof. When your agent automates a sales flow, you need a screenshot of the final state. When it processes an invoice, you need a PDF of what it saw. When it tests a checkout, you need a video of what happened.
2. Compliance and audit trails. If you're handling customer data or regulated workflows, regulators ask for evidence. "The agent accessed this page. Here's a screenshot. The agent submitted this form. Here's a video proving it."
3. Debugging and learning. When an agent fails, you need to see what it saw. Screenshots show you the DOM state when the click failed. Videos show you the interaction sequence. PDFs let you archive what the agent was processing.
WebMCP handles automation. It doesn't handle documentation.
The hosting question
You can run Puppeteer MCP or browser-use locally and get automation. But you still need:
- Screenshots (requires taking them somehow)
- Video recording (requires recording infrastructure)
- PDF generation (requires Chromium, which breaks in serverless)
- Rate limiting and audit logging (requires infrastructure)
Hosting this yourself means managing all of it.
The PageBolt model
PageBolt is the documentation layer. Your agent calls it when it needs proof of what happened:
Agent workflow:
1. Navigate to page (native MCP)
2. Fill form (native MCP)
3. Click submit (native MCP)
4. take_screenshot() → PageBolt API → get PNG proof
5. record_video(steps) → PageBolt API → get MP4 audit trail
6. generate_pdf() → PageBolt API → get archived PDF
The agent does the automation. PageBolt creates the audit trail.
Security data point
5,877 out of 10,631 MCP tools score poorly on security (CWE coverage). The most common vulnerabilities: no rate limiting, no audit logging, direct filesystem access.
PageBolt's model fixes this by design:
- Rate limited: 10–100 calls/min per API key. Brute force attacks fail instantly.
- Audited: Every call logged with timestamp, user, action, result. Compliance teams can query it.
- Scoped: Agent never gets filesystem access. Never gets raw browser access. Just API calls.
Real example
Scenario: Your AI agent processes customer refund requests. Compliance audit happens. They ask: "Show me what happened on March 2."
Without visual documentation:
Agent ran process_refund()
Agent navigated to /refunds
Agent clicked submit
[end of logs]
Compliance says: "That's not a proof of anything."
With visual documentation:
Agent ran process_refund()
Agent navigated to /refunds
Agent clicked submit
Screenshot: /audit/2026-03-02-refund-1.png (shows confirmation page)
Video: /audit/2026-03-02-refund-1.mp4 (shows entire flow)
PDF: /audit/2026-03-02-refund-1.pdf (archived state)
Compliance says: "That's auditable."
The distinction
- WebMCP / native automation: Agent can do things
- PageBolt: Agent can do things and prove it happened
One is a capability. The other is accountability.
For production AI workflows, you need both.
Getting started
PageBolt integrates with any MCP-compatible agent. Call it when you need visual proof.
Free tier: 100 requests/month. Enough to audit 20–30 complex workflows per month.
Get started at https://pagebolt.dev
This article reflects a positioning shift: PageBolt is not a Puppeteer replacement or a WebMCP competitor. It's the audit and documentation layer that sits alongside any browser automation tooling—native, MCP, or self-hosted.
Top comments (0)