The fast path: “just coding”
Ask the agent, in plain language:theo_setup_key with use_case="coding". Theo:
- Creates a new API key.
- Pins it to a single strong code engine as a deterministic passthrough — no intent classification, skills, or guardrail routing, so every request goes straight to the coding model with no surprises.
- Returns the raw key (shown once) and a dashboard link.
Use cases
Pass a
model to override the default for any use case, e.g. use_case="raw", model="anthropic/claude-opus-4.8".
Route by load (or any per-mode split)
Want light requests on a fast engine and heavy ones on a reasoning engine? Describe it, and the agent passes arouting array — Theo’s classifier splits each request into a mode and the binding pins that mode’s engine. No graph to build.
theo_setup_key with:
fast, think, code, research, roast, genui, data_extraction. Engine ids include arca-velox-5.1, arca-magnus-5.1, anthropic/claude-sonnet-4.6, anthropic/claude-opus-4.8, anthropic/claude-haiku-4.5, google/gemini-2.5-pro, openai/gpt-4.1.
Custom logic → a full graph
For arbitrary branching/guardrail logic the user describes in prose, the agent can pass agraph ({ enabled, nodes, edges }) — a full orchestrator graph, applied as-is. See the Canvas graph schema for node/edge shapes. graph takes precedence over routing and use_case.
Configure a key you already have
theo_configure_key. Same presets, applied to the key you name.
It round-trips to the dashboard
Both tools return a link like:Requirements
- The MCP server must be authenticated with a key that has the
billingscope (needed to create keys). See Install. - The created key ships with
completions,skills, andtoolsscopes.
See also
- MCP tools reference — full input schemas for
theo_setup_key/theo_configure_key. - Orchestrator guide — the dashboard surface the link opens.
