mode: "echo") makes no model
call and has no platform fee or model cost. Read
Model access and billing for pricing and who bills each charge.
Millwork is for developers whose application uses more than one model or
provider and needs to know what ran and what it cost. Start with
Run a model, or read on for one run from start to
finish.
Govern your AI workflow end to end
As your application grows, models multiply, providers change, and agents join the work. Millwork brings those choices together in an AI strategy envelope: you decide what may run, set the rules for each task, and use its receipt to inform the next decision.- Decide which models and kinds of data your application may use.
- Size each task with a model-usage budget and runtime limit. Read how the budget works.
- Run it on a ready saved model allowed by your data rules.
- Measure what happened. Inspect the selected model, recorded usage, and output check.
- Improve with evidence. Change a model, a limit, or a check, and compare receipts.
One run, start to finish
In the API and dashboard you will also see the name Solver, the part of Millwork that runs your task. This diagram follows a successful live run from your first request to its result and receipt.1
Make a model available
Add a model from your organization’s catalog, which can include a model
Millwork provides, or connect a provider
account your organization already uses. The API calls a saved model an
arm.2
Send the task and your limits
Describe the outcome you want. Set the data rules, model-usage budget, and
runtime limit. Send a request key in the
Idempotency-Key header so a retry does not start the work twice. The API
calls a run an execution.3
Millwork picks a model and runs it
Millwork keeps the saved models that fit your limits and picks one. The
run does not finish inside your request: read its status until it is
final.
4
Millwork checks the output
A basic check confirms only that output is present.
It does not judge quality. You can attach a stricter check of your own
(the API calls it a verifier,
verifier_id). See
Connect an output check.5
Read the result and the receipt
Use the result in your application. Keep the receipt as the record of
the selected model, recorded usage, and checks. It does not contain your
prompt or output.
Results and receipts
Use the result. Keep the receipt. A completed live run has both. Test, failed, cancelled, and expired runs have a receipt only. For a completed live run on a catalog model, the records also identify the model requested, the model the provider reported, and the provider connection. Results and receipts explains the details available for each kind of saved model and how long a result stays available. Today the private preview starts with text models. You can also save an agent endpoint or a reusable skill as an option where your organization has one; the same run and receipt apply. Two more things are planned: routing that learns from your past runs, and agents that Millwork runs for you. See What is available today.Where to start
See how a run works
Follow one task from your limits to its result and receipt.
Run a model
Submit one live task with public or test data and read its receipt.
Evaluate fit
Compare the preview with your requirements.
Private-preview access is available to approved organizations.
Contact Millwork to apply.
See what is available today.