Skip to content
SceneRecap

Skills, exports and MCP for agents

A recording that has been analysed already holds the thing an agent needs: the steps somebody took, in order, each tied to the second it happened. These three exports hand that to an agent in the shape it can use, and none of them invents a step that was not recorded.

Export a flow as a skill

On a ready recording, open Export and choose Agent skill (.zip). The zip holds one folder: a SKILL.md and up to twelve step screenshots under references/. It is the open Agent Skills format, which Claude Code, claude.ai, the Claude API and Codex all read, so one export works everywhere.

Where the folder goes:

  • Claude Code: unzip into .claude/skills/ in a project, or ~/.claude/skills/ for yourself.
  • claude.ai: upload the zip itself under Settings, Features (needs a plan with code execution).
  • Codex: unzip into .agents/skills/ in the repo, or ~/.agents/skills/ for yourself.

What is in the file:

  • A one sentence description saying what the flow does and when to use it. That is the only thing an agent reads before it decides to open the skill, so it is written for that job.
  • The goal, and what the flow assumes before step one.
  • The numbered steps, taken from the recording’s step-by-step guide word for word, each with the result to expect, the timestamp it happened at, and a screenshot where there is one.
  • How to tell it worked, and what to watch out for, both drawn from the recording.
  • A recovery section telling the agent to find controls by their label, to stop and ask before anything that submits, pays, sends, publishes or deletes, and never to invent a step.

What a skill is not

  • It is a snapshot of one recording of one version of an app. When the UI moves, the skill does not: export it again from a newer recording.
  • It contains no code. We never put executable scripts in a bundle generated from somebody’s screen recording.
  • A skill needs a step list, so a recording analysed in another mode has a step-by-step guide written first. That costs one AI analysis from your monthly allowance, and the export menu says so before you click.
  • Nothing from a sensitive field is in there. Typed text is never captured in the first place, and the export keeps the same privacy rules as every other export: paths without query strings, no keystrokes.

Skill export is on Pro and up.

A bug report written for a coding agent

On a Bug report recording, For coding agents (.md) downloads the report in the order an agent wants it: what is broken in one line, whether SceneRecap could reproduce it and how often, the shortest path that still breaks it, the step it fails on, the requests and console lines with their timestamps, the environment, and a closing instruction to reproduce before fixing. There is a Copy for Claude Code / Codex item in the same menu when you would rather paste it.

If the Repro Agent has run, its verdict is quoted as it stands, with the rate (three out of three runs, two out of three) and the line that bounds it: reproduction drives the recorded flow headlessly against an origin you control, so it reflects that environment and reports when it could not reproduce. If no run has been made, the report says that it has not, rather than implying the steps were confirmed.

Connect over MCP

SceneRecap speaks the Model Context Protocol, so an agent can reach your recordings itself instead of waiting for you to export one. Six tools:

  • search_recordings: find a recording by what happened in it.
  • get_recording: the analysis, with timestamps, and a link to watch it.
  • get_bug_report_for_agent: the report above.
  • export_skill: the SKILL.md text, the screenshot paths and a download link.
  • start_repro_run and get_repro_result: ask for a reproduction run against an origin you control, then read the verdict.

The setup snippets for Claude Code, Claude desktop and Codex are on your Developers page, next to the API keys. The server runs locally and talks to our API over HTTPS with the key you give it.

It is not published to npm yet, so the snippets run it from a checkout of the repository rather than through npx. We would rather ask you to clone something than hand you a command that fetches a package name we do not own, with your API key in its environment.

What the server will and will not do

  • It can reach exactly what the API key can reach: one workspace, one plan, the same monthly allowances. It has no database of its own and no privileged path.
  • It authenticates with a SceneRecap API key, not with OAuth. That makes it right for your own agents and not something to hand to a third-party service.
  • A reproduction run costs one agent run from your monthly allowance, only runs against an origin your workspace has proven it controls, and never signs in: a login wall ends a run as blocked.
  • Tool calls are logged as tool calls, by name. Arguments, results and recording contents are not sent anywhere.