Theo is not a model proxy. Every API call passes through a multi-stage pipeline that handles classification, skill injection, model selection, tool execution, and response formatting.Documentation Index
Fetch the complete documentation index at: https://docs.hitheo.ai/llms.txt
Use this file to discover all available pages before exploring further.
The Five Stages
1. Intent Classification
Determines the task type (fast, think, code, image, research, roast, etc.) using a multi-tier classification system. Skipped if you set mode explicitly.
2. Skill Loading
Loads domain expertise from two sources: the user’s installed skills (persistent) and per-request skills (ephemeralskills[] array). Each skill injects system prompt extensions and tool definitions.
3. Engine Routing
Selects the optimal Theo engine based on the resolved mode. If the primary engine is unavailable, failover routes to the backup automatically.4. Agent Loop
For requests with tools, Theo enters an iterative loop: the engine reasons, calls a tool, observes the result, and continues until done. For simple prompts with no tools, this collapses to a single call.5. Response
The final output is formatted with content, engine metadata, tool execution log, artifacts, follow-up suggestions, and usage data. The request is billed, cached, and audited.Performance
| Scenario | Typical Latency |
|---|---|
| Fast mode, no tools | ~500ms |
| Think/code mode, no tools | 1-3s |
| Agentic with tool calls | 2-5s |
| Image generation | 5-15s |
| Research (async) | 30-120s |