- Can my application run a task through Millwork?
- Can I set limits and confirm which model and provider answered?
- Does the private preview fit my data and operating requirements?
mode: "echo"), which does not call a model, so it has no platform fee and no
model cost. Each live run in step 3 has a platform fee when Millwork accepts it,
plus model usage. Read Model access and billing before
step 3.
Before you begin
- Your Millwork organization is approved for the private preview.
- You created an organization API key.
- You use only public information or test data created for this evaluation.
- Your model catalog lists a model you can use, or you can connect a provider account.
1
Check your access
Complete Make your first API call.Pass when: the test run’s
status is completed, its receipt’s mode is
echo, and selected_arm is null.2
Choose how you reach a model
Use a model already listed in your catalog, or
use your provider key to connect a provider account
your organization already pays for.Read your catalog with
GET /v1/model-catalog, or open the dashboard’s model
list.Pass when: your catalog lists a model you can save for runs.3
Run one public task
Complete Run a model. This step has a platform fee and
model usage.Pass when: the run’s
status is completed, a result exists, and a
receipt exists.4
Check which model and provider answered
For a run on a catalog model, compare the model Millwork requested with the
model the provider reported. Check the provider, who bills the model usage,
and the recorded usage.Read the four facts from the result you saved in the previous step:
billed_by is customer when your provider key was used and millwork when
Millwork provided the model. The receipt carries the same names under
slices[0].model_provenance and model usage under totals.usd. The platform
fee is its own line, totals.platform_fee_usd. To investigate individual
attempts, read the run’s events at
GET /v1/executions/{executionId}/events.If model_provenance is null, Millwork selected a model saved directly with
a provider model ID. That run does not prove which model the provider
actually ran. Record this gap if provider-reported identity is a requirement;
How model routing works explains the distinction.Pass when: the result and the receipt name the same model and provider.5
Check your request and records
Confirm that:
- the request contained no provider key;
- the task included data, cost, and time limits;
- the receipt did not copy the task or the model output.
Evidence checklist
Decide what comes next
Connect an output check
Host a handler for an acceptance check you already run.
Continue building
Add safe retries, cancellation, key replacement, and receipt exports.
Compare another provider
Connect another account and compare the models it adds to your catalog.
Check what is available
See what the private preview includes today and what is planned.
Ask a question
Discuss a requirement without sending credentials or customer data.