A tool that just waits sounds trivial until an agent is polling something that takes a minute to become ready — models are bad at pausing and good at retrying immediately. The part worth noticing is the configured maximum: without it, a model that asks for an hour gets an hour. With it, an over-long request comes back as a refusal that names the limit, so the workflow keeps moving instead of stalling somewhere you cannot see.
An MCP server whose whole purpose is to make an agent wait. It exposes a single tool that pauses for a specified duration and then tells the model it can continue, which gives a workflow a way to sequence steps around something that needs time to finish.
- Execution paused for a number of seconds, after which the model is told it can carry on — `sleep`
- A ceiling you control: ask for longer than the configured maximum and the tool refuses with a message naming the limit, instead of hanging
- A default maximum wait of 60 seconds, overridden with `MCP_SLEEP_TIMEOUT`
No account and no key. uv installed; the client entry runs `uvx mcp-sleep`, with `MCP_SLEEP_TIMEOUT` in the env block if you want a different ceiling. `uvx mcp-sleep --transport sse` runs it over SSE for clients that prefer that, and a Docker image is available as an alternative command. Also installable through Smithery or mcp-get. The pyproject name is `mcp-sleep` (0.1.1).
One command — uvx mcp-sleep