Skip to content

The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.

5.7 KB
Raw Blame

description: >- One subscription, many models. Each task picks its model with a hint: reasoning goes to a strong model, quick work to a fast one, images to a vision model. icon: route

Automatic model routing

Different parts of an agent want different models. Long reasoning wants a frontier model. A quick "fix this typo" wants a fast, cheap one. Images want a vision model. OpenHuman has a built-in router that picks for you.

How a request is routed

The model parameter on a chat call takes one of two forms:

  • A concrete model name, such as anthropic/claude-sonnet-4. It goes to the default provider with that exact model.
  • A hint prefix, such as hint:reasoning. The router looks the hint up in the route table and resolves it to a provider and model.

That is the whole rule. The router strips a hint: prefix, looks up the rest in the route table, and falls through to the default provider with the name unchanged if there is no entry. A name without the prefix is never rewritten.

The route table is [[model_routes]] in config.toml, with one entry per hint. [[embedding_routes]] does the same for embeddings. You can remap hints at runtime without restarting the core.

Common hints

Hint Typical target When it is used
hint:reasoning A strong reasoning model Multi-step planning, math, code-heavy turns
hint:fast A fast, cheap model UI helpers, autocomplete, small classification calls
hint:vision A vision-capable model Screenshots, image attachments, OCR
hint:summarize A model good at compression Background summarization
hint:code A code-tuned model Native coder turns
hint:burst A high-throughput, low-cost model Cheap, latency-tolerant work for high-fanout agents

The mappings are configurable, and the defaults ship with sensible per-provider routes.

One subscription, or your own

By default, routing happens behind a single OpenHuman subscription. You don't hold separate API keys for Anthropic, OpenAI or Google. The backend brokers access and the router picks the right model for each task.

The subscription is a default, not a requirement. The same router works with your own provider key or a local model on a runtime you run yourself (Ollama, LM Studio, MLX, or any OpenAI-compatible server). You choose per workload and can mix all three. See Local models and bring your own key for setup and for what each route supports for chat, vision and embeddings.

Overriding routes

  • Globally: [[model_routes]] in config.toml sets the route table at startup.
  • Per call: pass a concrete model name with no hint: prefix. The router falls through to the default provider with that exact model.
  • For a skill: a skill can pin a hint or a model in its manifest.
  • Per agent: an agent definition can pin its own provider and model, which wins over the table.

Default model

Settings > Connections > LLM > Routing has a Default model row. It sets a model from the managed catalog that every managed chat turn runs on, instead of the anonymous chat tier. It opens on DeepSeek V4 Flash. The model pill in the composer can still override it for one conversation. The specialized tiers (reasoning, coding, vision and summarization) keep their own routing. The rows below it send each workload to Managed, a BYOK provider, a local runtime or Claude Code.

Per-agent model pins

Sub-agents can pin an exact model without turning off automatic routing for the rest of the app. This helps when an orchestrator or team lead needs a stronger model while high-volume leaf agents stay on a cheaper one.

An inline pin wins for a single delegation:

{
  "agent_id": "presentation_agent",
  "model": "anthropic/claude-sonnet-4",
  "prompt": "Build a five-slide deck from the Q3 report."
}

Lasting defaults go in config.toml:

[orchestrator]
model = "anthropic/claude-sonnet-4"

[teams.planner]
lead_model = "openai/gpt-5.1"
agent_model = "groq/llama-3.1-8b-instant"

[teams.image]
agent_model = "openai/gpt-5.1"

The router picks the model in this order:

  1. The inline model on spawn_subagent or an archetype delegation call.
  2. [orchestrator].model, or [teams.<agent_id>], or the alias without _agent ([teams.image] for image_agent).
  3. The archetype's own model hint and the normal route table.

For [teams.*], lead_model applies to agents that can delegate, and agent_model applies to leaf workers. If only one is set, it is used for both roles.

Why this is more than a model switcher

Routing is not a dropdown. The agent loop itself picks a hint based on what it is about to do. You don't pick the model, the task does. That is the difference between multi-model and smart routing.

See also