Skip to content
The Model Company

Live · Agent skill

Model-building, as a tool call.

Paste this into your agent

and it sets up your model with you.

Free in the preview, where jobs are simulated. The API, MCP server and CLI are planned.

Describe the job. Your agent does the setup.

  1. Step 1: Shape the task

    task + labels

    Agreed with you, in one sentence the builder can read.

  2. Step 2: Prepare examples

    examples.csv

    Built from your data and checked against the builder's rules.

  3. Step 3: Open the builder

    /create?task=…

    Task and labels filled in. You upload the file and start.

Live by invite · Shadow mode

Shadow first. Swap when it wins.

Your new model answers a free copy of every live call, next to the model you use today. Both are graded on what really happened. When yours is ahead, swap them, and swap back in one click if you ever need to.

for shadow calls while your model proves itself
$0
from your examples to a safe swap*
Same day
to swap back, whenever you want
1 click

*A liberal estimate for the sample support ticket router: up to 23 minutes to train, then about 10 hours in the shadow to grade 1,000 calls at around 100 an hour. Quieter workloads take a few days.

Illustrativeshadow calls · $0.00
  1. 0:00training · 2,400 examples · 6 labels
  2. 0:23ready · 92.1% on 240 held-out tickets
  3. 0:23shadow on · every live call is copied to your model
  4. 2:23200 graded · yours 91.0% · incumbent 86.5%
  5. 6:23600 graded · yours 91.7% · incumbent 86.2%
  6. 10:231,000 graded · yours 92.3% · incumbent 85.8%
  7. 10:23p95 13 ms vs 2,441 ms · $0.036 vs $3.24 per 1k calls
  8. 10:23✓ ahead on quality, speed and cost · clears your bar · ready to swap
  9. 10:24swapped · yours answers live · incumbent on standby

Three ways to build.

Read the full reference

Hand the setup to your agent.

Free in the preview. No sign-up, no API key.

Questions.