Skip to content

MCP — AI assistants

MCP lets an AI assistant read your fleet and the market, and — on the top plan — act on it. The design assumption is that the assistant will occasionally be wrong, so every action it can take has a ceiling you did not have to think of.

Reading works on Professional · Optimize and above: fleet, servers, telemetry, market, earnings, host requirements. An assistant with read access can answer “which of my machines is underpriced” and cannot do anything about it.

For most people this is the whole feature.

Writing requires Professional · Automate, and every write is bounded:

Limit Ceiling
Price change in one move 15%
Reprices 20 per hour
Writes 60 per hour
Market moves 10 per day
Reboots 5 per day
Power limit 600 W
Concurrent tasks 5

These are ceilings, not targets. An assistant that tries to reprice 40 times in an hour is stopped at 20 — the limit does its work whether the reason was a bad prompt, a loop, or a genuinely volatile market.

Every tool carries a risk class. By default everything that changes something asks you first.

On Automate you may pre-approve the lowest risk classes, so routine reads and small moves stop interrupting you while destructive operations still require a human. Reboots, power-offs and anything that ends a rental never become silent.

Connect from the panel’s MCP section. Two ways in:

  • OAuth, for clients that support it — these appear in your list of connected clients and can be revoked individually.
  • A token in a header, for everything else. These do not appear in the client list, so keep track of what you issued.

Client and grant counts are capped per plan: 3 clients on Optimize, 10 on Automate.

Calls are capped per minute and per hour — 60/minute and 1,000/hour on Optimize, 120/minute and 3,000/hour on Automate. An assistant that hits them is told to slow down rather than being disconnected.