How ClawAI works
From creating an account to reading an answer, and everything in between: how a model gets chosen, what it is given to work with, and how your usage is counted.
Last reviewed 2026-07-25
The whole thing in eight steps
Each of these gets its own section below.
Create an account. an email address and a password. The free plan starts immediately, with no card required.
Pick a plan. free to start; paid plans from $5 a month raise your allowance and unlock the larger models and the multi-model modes.
Start a conversation. type a message, optionally attach files, and either choose a model or leave it on Auto.
ClawAI picks the model. the message is classified and routed to a model that handles that kind of work well.
Context is assembled. relevant memories, attached context packs and the parts of your files that matter are added to the prompt.
The answer streams back. text arrives as it is generated, with live timing and, where the model exposes it, its reasoning.
You see what happened. which model answered, why it was chosen, what it cost, and exactly what context it was given.
Usage is counted. the cost is metered against your allowance in cost-normalized tokens and shown on your usage page.
Creating an account and choosing a plan
Registration
An email address and a password is all it takes. There is no provider account to create, no API key to paste and nothing to install — ClawAI holds the provider relationships on your behalf.
The plans
Seven tiers, from free to Unlimited at $200 a month. Every paid plan reaches every model; what changes is how much you can use and which multi-model modes are unlocked.
- Free — $0. A small daily allowance, entry-tier models, and one trial run each of Compare, Judge and Research.
- Starter $5, Plus $10, Pro $20 a month. Rising allowances, with premium models and larger Compare and Judge quotas as you go up.
- Team $50, Scale $100 a month. Unlimited Compare, Judge, Critic and Research, plus many more workspace connections.
- Unlimited $200 a month. Unlimited conversations and messages, with a fair-use boundary on premium models that you are warned about before you reach it.
Pay yearly and you pay for ten months instead of twelve. You can change plan or cancel at any time — an upgrade applies immediately, a downgrade at the start of your next billing period.
The models you can reach
One subscription, every family below. You can switch between them inside a single conversation.
- Anthropic Claude
- Careful reasoning, long documents, and the most reliable code review in the roster.
- Claude Opus 5 · Claude Sonnet 5 · Claude Fable 5
- OpenAI GPT
- Broad general capability with strong tool use and structured output.
- GPT-5 · GPT-5 mini
- Google Gemini
- Huge context windows, fast responses, and native image, audio and video input.
- Gemini 3 Pro · Gemini 3 Flash
- Moonshot Kimi
- Long-context analysis and agentic work at low cost.
- Kimi K2
- Zhipu GLM
- Strong bilingual reasoning with an explicit thinking mode.
- GLM-5.1
- Alibaba Qwen
- Excellent code generation across a wide range of model sizes.
- Qwen3
- DeepSeek
- Mathematics, algorithms and step-by-step derivations.
- DeepSeek V3.2
- xAI Grok
- Fast conversational answers with awareness of recent events.
- Grok 4
- Amazon Bedrock
- AWS-hosted models for teams already standardised on Amazon.
- Nova Pro · Bedrock-hosted Claude
Which models a plan can reach depends on the tier: entry models on Free, everything including the largest frontier models from Pro upward. New models are added as providers release them.
How ClawAI decides which model answers
In Auto mode every message is classified before it is sent. These are the classes that change the outcome most:
- Code
- writing, reviewing, refactoring or debugging code goes to the models with the strongest coding track record.
- Hard reasoning
- multi-step problems, proofs, architecture decisions, and anything where being wrong is expensive.
- Writing
- drafting, editing and rewriting, where tone and readability matter more than raw reasoning depth.
- Data and documents
- long inputs, spreadsheets and reports, routed to the models with the largest context windows.
- Images
- requests to generate or interpret an image go to a model that can actually handle pixels.
- Everyday questions
- short factual questions and quick edits go to a fast, inexpensive model — most messages land here.
You can always override it
Pick a model from the selector and it is used for that conversation until you change it. You can also bias Auto toward speed, reasoning depth or cost, which keeps the per-message classification but changes what it optimises for. Whatever you choose, the answer tells you which model produced it.
Using more than one model at once
For work where one answer is not enough, paid plans unlock modes that put several models on the same prompt.
- Compare
- one prompt to up to five models, answers side by side, with latency and token counts for each.
- Consensus
- several models answer and ClawAI synthesises one response from where they agree, while flagging where they do not.
- Escalation
- start cheap and fast, and move up to a stronger model automatically only when the answer is not good enough.
- Best-of-N
- generate several candidates and keep the highest-scoring one.
- Judge and Critic
- an independent model scores the answer against explicit criteria, and a Critic pass writes out what is weak and why.
Free gives you one trial run each of Compare, Judge and Research. Paid plans have monthly quotas that rise with the tier, and Team and above are uncapped. Repair, Verify, Role packs and Pipelines round out the set on the higher plans.
What the model is given besides your message
Three layers of durable context can be assembled into a prompt, all of them under your control.
- Memory
- facts, preferences and instructions worth carrying between conversations. Candidates go to an approval queue rather than being saved silently.
- Context packs
- named bundles of reusable text, files, links and memory references that you attach to a conversation when they are relevant.
- Files
- uploaded documents are chunked and indexed, so only the passages relevant to your question are pulled in.
Memory and context are per-conversation switches. When they are on, the answer carries a receipt listing exactly which memories, pack items and file chunks were used, and how much of the token budget each one took.
What every answer tells you
Routing you cannot inspect is just a black box with better marketing. Every answer carries its own record.
- The model
- the provider and exact model version that produced the answer, including when a fallback stepped in for the first choice.
- Why that model
- the classification, the confidence score and the reasons behind the routing decision.
- What it cost
- input and output tokens, latency, and the allowance drawn down in cost-normalized terms.
- What it saw
- the memories, context pack items and file chunks assembled into the prompt, in the order they appeared.
- Where you stand
- your usage page aggregates all of it into daily and monthly balances, broken down by model.
All of this is visible the moment you read the answer, not in a report you have to request.
How usage is counted
One number covers every model, and it stays fair across models whose prices differ by more than an order of magnitude.
Cost-normalized tokens
A token from an expensive frontier model draws down more of your allowance than a token from a cheap, fast one, in proportion to what it actually costs. That way a single allowance figure works whichever model you use, and you are never penalised for picking the right tool for the job.
- Light models
- fast, inexpensive models draw down slowly. A day of short questions barely moves the number.
- Mid-tier models
- the general-purpose workhorses sit in the middle. This is where most sustained work lands.
- Frontier models
- the largest reasoning models draw down fastest. Worth it for hard problems, wasteful for small talk — which is exactly what Auto routing is for.
The windows
- Daily
- resets every 24 hours, so one heavy day never spoils the rest of the month.
- Weekly
- smooths out bursts on the higher tiers, where a daily limit on its own would be too coarse.
- Monthly
- the headline figure on your plan, resetting with your billing period.
You can see the balance on all three windows at any time. When you reach one, ClawAI tells you which limit it was, how much is left on the others and when it resets.
On the Unlimited plan, conversations and messages genuinely have no cap. Premium frontier models carry a fair-use boundary so a single account cannot run up unbounded provider cost — you are told clearly as you approach it, never cut off without warning.
For organisations
If your organisation cannot send data to a third-party model provider, ClawAI can be deployed inside your own network running open-weight models on your own hardware. Nothing leaves your infrastructure. It is scoped per engagement, not sold as a plan.
Start on the free plan
A minute to sign up, no card required. Ask a few questions, run a Compare, and see whether the routing earns its keep.