Get started

Connect one workflow. Test the next change.

Use Stimulir to route a workflow, simulate cases, compare improvements and control what runs next.

Updated

These docs are for teams implementing Stimulir. For deployment options and support, start with Enterprise. You can bring a business process even if you have no AI system yet.

1. Route a workflow

Open Console, select your organisation and project, then choose an available model in Models. Connect workflow steps in Agents, or use the API example below for an existing application.

Create an application key in Settings → Project → API keys. Keep it in your application's environment. Copy the project ID from Settings → Project → General.

bash
pip install stimulir
export STIMULIR_API_KEY='hyb_...'
export STIMULIR_PROJECT_ID='your-project-id'
export STIMULIR_MODEL='your-enabled-model-id'

Send a request

python
import os
from stimulir import StimulirClient

client = StimulirClient(tags=["first-request"])
reply = client.agent(
    prompt="Reply with one sentence describing a useful AI workflow.",
    model=os.environ["STIMULIR_MODEL"],
)
if reply.error:
    raise RuntimeError(reply.error)
print(reply.content)

In Traces, filter to first-request in the same project. Check the response, status and usage. Calls can consume provider credit.

2. Simulate with real cases

Collect representative inputs and expected outcomes. In Lab → New evaluation, select the dataset, current version and candidate. Use the same cases and criteria for the comparison.

Simulate and review a change →

3. Improve and compare

Inspect failures, test a prompt or model change, and compare quality, cost and speed. Review the result before changing production.

Run a controlled improvement →

4. Automate with controls

Define when checks run, their budget and who approves a production change. Start with one bounded run before repeating the process.

Set the operating rules →

If you get stuck

Check the project, model route and error code. Use the Console guide or request troubleshooting. Never share credentials or private request content in support messages.