MCP tools
The same catalogue, as tools an agent calls directly. This is what the install command connects you to.
Connecting
bash
claude mcp add --scope user --transport http sequel https://sequelmcp.com/mcpNo key in the URL. The client discovers the authorization server, opens a browser, and you approve there — the token goes to the client rather than through your clipboard.
The tools
ToolTakesReturns
search_modelsquery, type, limitMatching models with price and latency
get_model_schemasmodelJSON Schema, capabilities, rates
run_modelsmodel + input, or runs/promptsTask, or tasks in order
get_upload_urlcontent_type, extensionPUT target and read URL
list_tasksstatus, models, datesRecent runs for your organisation
Batching
run_models takes several at once. Prefer prompts — it is a fraction of the JSON of a runs array, and a long one gets truncated mid-string by smaller models.
json
{ "model": "black-forest-labs/flux-1-schnell", "input": { "steps": 8 }, "prompts": ["a ceramic mug", "a leather wallet", "sunglasses"] }
action_info
Every tool takes an optional action_info: a short phrase saying what the call is for. It is display-only and stripped before the call runs, but it is what turns a trace of tool names into something a person can read.