Your agent sets itself upYour agent sets itself up

A harness that speaks MCP connects over OAuth: the harness offers a Connect button, your browser opens a consent page naming the account it would publish to, and the key it holds is revocable from your roster.

Everything else takes a key: copy it from the API keys page and set SUPER_ARTIFACTS_KEY in the environment the agent runs in. The skill it installs carries no secret at all, which is what makes it safe to commit, sync between machines, or publish to a registry.

Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, Hermes, OpenClaw, Zed, Windsurf, Devin — each has its own path, because each harness persists instructions and keys its own way. The Agents page in the dashboard lists the exact commands, checked against each tool’s own documentation.

The skill at /skill/SKILL.md carries no secret at all — it reads the key from the environment — which is what makes it safe to commit, sync between machines, or publish to a registry.

A hosted chat assistant cannot make an arbitrary HTTP request; it reaches external tools through a connector. So there is an MCP endpoint that wraps the same deploy call, authenticated by the same revocable deploy key — and the endpoint speaks OAuth itself, so the client’s Connect button does the whole handshake.

Reading collected data back over MCP is built: super_data returns an artifact’s rows, and super_data_query returns totals and distributions. See MCP.