Model routing and orchestration in ClawAI

Seven routing modes, not one hidden default

ClawAI classifies each message and can send it to a fitting model automatically, or you can decide the policy yourself. The routing modes are Auto (classifies by task and picks a strong model for that class), Manual Model (pins one model to the conversation), Local-Only (every request stays on hardware you control, via Ollama or llama.cpp), Privacy-First (a separate mode with its own priorities for keeping a request out of a general-purpose cloud path), Low Latency (prefers the model that can respond most quickly), High Reasoning (prefers the strongest reasoning model regardless of speed or cost) and Cost Saver (prefers the cheapest model that can still handle the request). See what is AI model routing, linked below, for how a router decides in general.

Nine ways to put more than one model on a problem

When one model is not enough, ClawAI’s orchestration primitives — recorded on the ledger under the ORCHESTRATION surface, separate from ordinary chat — are Compare (up to five models on one prompt, side by side), Consensus (synthesise one answer from where several models agree, and flag where they do not), Escalation (start cheap, move up automatically only when quality falls short), Best-of-N (generate several candidates, keep the strongest), Repair (fix a specific defect in an existing answer instead of regenerating it), Verify (a second model checks correctness with a configurable revision limit), Role packs (a small team of role-specialised models handing off to each other), Pipelines (chain several of these stages into one named, re-runnable workflow) and Judge and Critic (an independent model scores a response against explicit criteria, with a written Critic pass on what is weak). Compare and Judge are each individually plan-gated (COMPARE_MODE, JUDGE_MODE, CRITIC_REVIEW); see what is AI consensus and what is an AI judge, both linked below, for how the assessment itself works.

What happens when a provider fails mid-request

A routing decision is not a one-time bet: if the provider or model a request was sent to fails mid-request, ClawAI can fail over to another model automatically, and the response records which model actually stepped in — not just the one originally chosen. See what is model fallback, linked below, for how that failover decision itself gets made.

Questions people ask

How many routing modes does ClawAI have?

Seven: Auto, Manual Model, Local-Only, Privacy-First, Low Latency, High Reasoning and Cost Saver. Auto is the default; the other six exist for when you want the routing decision yourself, or biased in a specific direction.

What is the difference between Compare and Consensus?

Compare shows every model’s answer to the same prompt side by side, with per-model latency and token counts, and leaves the reading to you. Consensus synthesises one answer from where the models agree and flags where they do not.

Can I see which model actually answered, and why?

Yes — every answer carries the provider and model that produced it, the reasoning behind the routing choice, and what it cost against your allowance. If a provider failed and another model stepped in, that is recorded too.