v2.4.0
Parallel workers are live →
Evals that run on every commit.
Tessel tests your LLM app against real conversations before it ships. One config file, one CLI, results right in the pull request.
Free up to 10,000 runs a month · no card · SOC 2 Type II
~/checkout-agent — zsh
Copy
curl
node
python
cli
1
2
3
4
5
6
7
8
9
$ curl https://api.tessel.dev/v1/runs \
-H "Authorization: Bearer $TESSEL_KEY" \
-d suite=checkout-agent -d model=gpt-5-mini
{
"id": "run_8f2k1",
"passed": 182, "failed": 4,
"p95_ms": 412
}
$
api operational · p95 412ms
sdk v2.4.0
//
02
why it exists
Prompts change every week. Your users notice every regression. Tessel turns each prompt edit into a test run, a diff and a red or green check, before anyone merges.
//
03
under the hood
Built for the pipeline you already have.
~/checkout-agent — zsh
Copy
curl
node
python
cli
1
2
3
4
5
6
7
8
9
$ npx tessel run --suite checkout-agent
✓ 182 passed ✗ 4 failed 41.2s
✗ refund_over_limit expected "deny"
✗ address_fr_cedex timeout 8000ms
→ report https://tessel.dev/r/8f2k1
$
api operational · p95 412ms
sdk v2.4.0
One command locally, the same check in CI.
//
06
pricing
Pay for runs, not seats.
//
07
questions
Before you wire it into CI.
What counts as a run?
One case, evaluated once, against one model. A 200-case suite tested on two models is 400 runs.
Do you train on our data?
No. Inputs and outputs are stored only for your retention window, encrypted at rest, and never used to train anything.
Can we run it without sending data out?
Yes. Team and Enterprise support self-hosted runners. Only pass/fail counts leave your network.