Point a Claude session at this file and ask it to use
skills/claude-computer-use/SKILL.md
for your task. For example:
Read
info/claude-computer-use.mdin the Machine Control checkout and follow its skill to use Computer Use for this TextEdit task. Prepare the scoped MCP setup if it is missing, and tell me whether this session needs restarting.
For a YA setup session on a dedicated disposable appliance, include that host context and the task's registration authorization. For example:
This is YA-hosted Claude on the selected dedicated disposable test appliance. Scoped host registration for this task is authorized on this appliance; preserve unrelated settings and leave model/AAR credentials alone. Follow the guide's skill to complete setup, then report the required fresh-session handoff. The coordinator will stop and verify this setup session before launching the task session; do not launch a nested agent.
The skill explains the tested YA user-scope host registration. A generated
MCP file or a claude --mcp-config suggestion alone does not load it into YA.
This task-local appliance mutation is cleaned up by clone disposal. It does
not authorize persistent user/global configuration changes on a personal
host; the CLI's per-session config remains supported for that case.
The skill includes a configuration helper and startup probe. Claude chooses each UI action; no Codex reasoning agent is involved. The genuine OpenAI runtime must already be installed and usable on the Mac that runs the MCP server. The checkout supplies instructions and an experimental adapter, not the proprietary runtime.
Current: YA-hosted Opus completed two TextEdit workflows through this adapter, and a subsequent fresh-session acceptance qualified agent-led skill setup followed by a separate task session. Any Claude host capable of stdio MCP and inline images is a candidate, but that is not qualification of every host, application or runtime version. Reading a file cannot itself add tools to a running session: use the host's supported MCP loading mechanism, or start a new session with the generated configuration. YA/AAR model setup remains separate.
For alternatives and product scope, see the competitive landscape. For measured results and limits, see Tactical 127.