Coding agents (MCP)

Playhead ships an MCP server, so the agent already working in your repo can explore the app, write a grounded spec, render it, and read the verdict.

Register it

// .mcp.json
{
  "mcpServers": {
    "playhead": { "command": "npx", "args": ["-p", "playhead-cli", "playhead", "mcp"] }
  }
}

The tools an agent gets

Tool	What it does
`playhead_explore`	validated locators for the current screen — the agent’s ground truth
`playhead_validate_spec`	schema and locator check, no render
`playhead_author`	draft a spec toward a goal
`playhead_render`	capture → compose → verify, returning the verdict with reasons and artifact paths
`playhead_jira`	attach the evidence and post a verdict comment to an issue

The loop

explore → write spec → render → read verdict → fix → re-render. The verdict is what makes this work unattended: the agent gets a machine-readable reason for every failure — this target was cropped, this click changed nothing, this frame was blank — instead of having to watch a video it cannot see.

Every locator the agent uses must come from an explore catalog. That single rule is the difference between a spec that renders and one that fails at capture time on invented element names.