Inspiration
When a browser call fails, developers often jump between application state, WebRTC statistics, developer tools, logs, and manual controls. This is slow, difficult to explain, and risky: an AI assistant may either lack access to the live runtime or receive too much authority to change it.
What CallScope does
CallScope is a browser-based WebRTC Rescue Room where a developer and an AI agent investigate and repair a live call together.
The workflow is:
Observe → Diagnose → Explain → Propose repair → Human approval → Apply → Verify
The developer starts a deterministic browser call and introduces a real fault by disabling the audio track. Through focused WebMCP tools, the agent inspects the actual runtime state, explains the evidence, stages a recovery plan, waits for explicit human approval, applies only the approved repair, verifies the before-and-after result, and generates a sanitized incident report.
Why WebMCP is essential
CallScope is not ordinary browser automation. The agent does not guess from screenshots or scrape UI text. WebMCP gives it structured, purpose-built access to the live technical state of the call.
This makes a previously difficult workflow practical: the developer and agent can inspect the same call, share evidence, agree on a bounded repair, and verify the result together—without giving the agent uncontrolled authority.
How it works
CallScope is built with Vue 3, Vite, JavaScript, and Pinia. It creates a deterministic browser-only WebRTC loopback, so the demo does not depend on a backend, PBX, SIP service, or external caller.
The WebMCP layer exposes narrow tools with structured schemas for inspection, diagnosis, recovery staging, approved repair, verification, and incident reporting. Read-only inspection is separated from state-changing recovery, and recovery remains blocked until the developer approves the staged plan.
Key features
- Deterministic browser-only WebRTC loopback lab
- Reproducible disabled-audio failure
- Structured WebMCP inspection and diagnosis tools
- Visible evidence and diagnosis timeline
- Staged recovery plan with explicit human approval
- Bounded repair through WebMCP
- Before-and-after recovery verification
- Sanitized incident report
- Lifecycle cleanup, error handling, and privacy-aware outputs
How we used AI and Codex
The AI agent acts as the diagnostic collaborator: it selects WebMCP tools, interprets structured WebRTC evidence, explains the likely cause, proposes a bounded recovery plan, and verifies whether the repair worked.
Codex helped translate the specification into an implementation sequence, challenge design assumptions, implement and review the application, harden approval and lifecycle behavior, design deterministic tests, improve privacy boundaries and error handling, and prepare the documentation and demo.
Try it
Open the live app in ChatGPT's in-app browser or Google Chrome with WebMCP enabled. Start the demo lab, disable the audio track, and ask the agent: What is wrong with my call? Follow the visible diagnosis, approve the staged recovery, and let the agent apply and verify the fix.
No account or credentials are required.
Limitations
CallScope demonstrates deterministic browser-side WebRTC diagnosis and recovery. It is not a backend monitoring platform or an infrastructure observability service. The focused demo centers on a reproducible disabled-audio failure so the human-agent workflow remains fast, reliable, and understandable.
Built With
- javascript
- pinia
- vite
- vue.js
- webmcp
- webrtc