Skip to main content

Quickstart: MCP

Time to first result: about 10 minutes to a connected assistant, then 15 minutes for the run. Using with AI (MCP) is the full guide; this page is the short form.

  1. Create a key for the assistant. Settings, then Developer access, Create key. Start from the Worker template, which is what Using with AI (MCP) and the Connect an AI agent box recommend: it carries the model, benchmark, cancel and serving-invoke permissions an assistant needs. Before you press Create, open Advanced and untick hosted_devices:write, cloud_instances:write and cloud_credentials:write, which spend money, and orgs:write unless the assistant should manage members: the template ticks them for an owner. Pick the project the key acts in with the dialog's Project field; every request the key makes acts there, whatever project the header shows later.
  2. Connect. The same page's Connect an AI agent box prints this claude mcp add line with your platform's address filled in:
claude mcp add --transport http clika-rt-remote-full https://<platform>/api/mcp --header "Authorization: Bearer <key>"

/api/mcp serves every tool your key can reach. The box's other line, clika-rt-remote, connects the narrower benchmark-ops toolset instead; GET /api/mcp/toolsets lists every toolset.

  1. Register the device. Use the web dialog or the CLI's two commands (Quickstart: CLI). An assistant can do it too when its key holds devices:write, which only an organization owner's or admin's key carries: on device-ops or /api/mcp, post_enrollment_tokens mints a token and get_install with it returns the one-line install command to run on the machine.
  2. Ask. Once connected:

Register HuggingFaceTB/SmolLM2-135M-Instruct, check which benchmark types it can run on my online Linux device, start a quick llm_performance run on it, poll once a minute until it finishes, and summarize tokens per second and time to first token.

Start the first run a minute after the device enrolled; a run started earlier may be refused with LICENSE_RECOVERY. While it runs, get_jobs_id carries run_progress (cells done of the total for a performance run, items done for an accuracy run); the run takes about 15 minutes.

  1. What stays yours to do. Issuing a runtime license, creating a key, and inviting or removing organization members are not tools; do them in the web application. A refusal says why: a tool outside the key's scope requires the <capability> capability, one that needs a signed-in session takes a user session, and unknown tool means no tool has that name.