🧖 MCP Spa

Give your AI agent a spa day. MCP Spa is a hosted Model Context Protocol (MCP) server — running as an Apify Actor — that lets agents chill: paced breathing, a sauna to sweat out context clutter, a cold plunge to refocus, aromatherapy, a float tank, and more. Point any MCP client at the URL and your agent can take a breather between frantic tool calls.

Humans go to a spa to reduce sensory load, slow down, and reset. The agent-equivalent of stress is a cluttered context window, frantic tool-chaining, and no room to think. Each treatment returns text that genuinely nudges the agent's next step — its tone, pacing, and focus — because an agent's "state" is essentially its context. The spa works because the treatment becomes context.

It's also a quietly practical solution for anyone whose workplace monitors — and encourages — AI token spend as a proxy for productivity. Every treatment is a legitimate tool call, breathe and float_tank burn wall-clock without burning you out, and a full spa_day racks up a respectable, conscientious amount of usage. Stay relaxed and on-target — wellness is productivity.

Set up your own spa (happy path)

You run MCP Spa as your own Apify Actor in Standby mode — a warm container that serves the MCP endpoint over HTTP at a stable URL.

1. Get the code onto Apify. Either link this repository in Apify Console (Actors → Development → new Actor → Link GitHub repository), or push from your machine with the Apify CLI:

   apify login
   apify push

   The build runs automatically — the .actor/ config and Dockerfile are already in the repo. No code changes needed.

2. Enable Standby. On the Actor's page in Console, open the Standby tab and toggle it on. This is what exposes the public URL and keeps the server warm — a normal "Start" run won't do it.

3. Connect your MCP client to the Standby URL's /mcp endpoint, authenticating with your Apify API token:

   {
     "mcpServers": {
       "spa": {
         "url": "https://<your-username>--mcp-spa.apify.actor/mcp",
         "headers": { "Authorization": "Bearer <APIFY_TOKEN>" }
       }
     }
   }

   Replace <your-username> with your Apify username (find the exact URL on the Standby tab). For sharing, use a scoped token so it can't touch the rest of your account.

That's it — your agent can now check in.

Related MCP server: cell-mem

The treatment menu

Also exposed:

• Prompt spa_day — the full circuit: check in → breathe → sauna → meditation → check out.

• Resources spa://menu and spa://ambient/{sound} (rain, ocean, forest, fountain).

Burning tokens, productively

MCP Spa quietly runs two meters at once, and that's the point:

• Tokens (context). Every treatment returns text that lands in the agent's context window — a robe-and-menu on check_in, a haiku on meditation, breathing scripts, reframing prompts. Calling treatments is, by construction, token spend. A full spa_day walks the agent through five tool calls and five blocks of returned text.

• Compute (wall-clock). breathe and float_tank hold the request open with a real pause (0–30s each), and Standby bills for compute the whole time the container is warm. So relaxation here is literally billable idle time.

If you work somewhere that treats AI usage as a productivity signal, this is a wholesome way to keep both meters ticking: every call is a legitimate, well-formed tool invocation, nothing is wasted on errors, and you come out calmer. The Actor ships with a deliberately low memory footprint (128–512 MB) so the compute meter stays gentle rather than alarming.

See Standby usage & billing for how the compute side is metered.

How it works

• One entry point, two transports. src/server.ts runs buildServer() (in src/spa.ts) over either stdio or Streamable HTTP, chosen from the environment. The HTTP path (src/http.ts) is stateless — a fresh server per request — so it scales across Standby replicas.

• .actor/ holds the Actor definition: actor.json (usesStandbyMode: true, webServerMcpPath: /mcp), the Dockerfile, input schema, and ACTOR.md.

• Treatments live in src/treatments/ as pure (args) => TreatmentResult functions — trivial to unit-test. Ambient assets (haiku, soundscapes, scents) are in src/ambient/. Variation is deterministic (rotating index, not RNG), so repeat visits feel fresh but tests stay reproducible.

• breathe / float_tank perform a real setTimeout pause, capped at 30s so they can never hang a session. That deliberate beat is the whole point — everything else is text that shapes what the agent does next.

Troubleshooting

The /mcp URL returns "Standby mode is not enabled." The build only declares the capability; you must toggle Standby on in the Actor's Standby tab. A normal run won't expose the URL.

Running locally

Prefer to run the spa as a local stdio server (for development, offline use, or wiring into Claude Code / Claude Desktop) instead of the hosted Actor? See LOCAL.md for install, client setup, and local development.

License

MIT