automl-agent-client

Python client and CLI for AutoML-Agent. Describe an ML task in one sentence, add your data, and five AI agents do the rest:

• plan the approach;
• search models and hyperparameters;
• write the pipeline and verify it;
• hand you a deployable script.

pip install automl-agent-client

1. Get an API key

Sign in at agentml.kitasid.dev. Open Settings → API keys, then create a key. It's shown once, so copy it. Then use either option:

agentml login                      # paste the key; it's saved for your user only
# or
export AUTOML_AGENT_API_KEY=amk_…

2. Python

from automl_agent_client import AutoMLClient

client = AutoMLClient()

run = client.runs.create(
    "Classify banana quality as Good or Bad. Accuracy must be at least 0.98.",
    files=["banana_quality.csv"],
)
for event in client.runs.stream(run.run_id):  # live, stage by stage
    print(event.stage, event.status, event.message)

record = client.runs.get(run.run_id)
print(record.status)  # completed
open("pipeline.py", "w").write(client.runs.code(run.run_id))

for chunk in client.runs.ask(run.run_id, "Why was this model chosen?"):
    print(chunk, end="")

Or in one call: record = client.runs.run("…", files=[…], timeout=900).

Async: the same methods, awaitable, on AsyncAutoMLClient:

async with AsyncAutoMLClient() as client:
    run = await client.runs.create("Forecast daily sales for 14 days", files=["sales.csv"])
    async for event in client.runs.stream(run.run_id):
        ...

Drift

Check fresh data against what a run learned from. If you include the target column, executed runs also re-score the trained model.

analysis = client.drift.analyze(run.run_id, "bananas_this_week.csv")
analysis = client.drift.wait(run.run_id, analysis.id)  # the model check finishes in the background
print(analysis.result.data.status)  # stable / moderate / drifted

Deployed models

Deploy an executed tabular run's trained model, then predict, explain and keep it fresh. Retraining reuses the run's pipeline and best hyperparameters on your new labelled data, with no LLM involved, and every retrained version waits for your approval.

dep = client.deployments.create(run.run_id, name="Banana quality")
client.deployments.wait_ready(dep.deployment.id)

out = client.deployments.predict(dep.deployment.id, [{"Size": -1.9, "Weight": 0.4, "Sweetness": 0.6}], explain=True)
print(out.predictions, out.probabilities, out.explanation.contributions)

client.deployments.feedback(dep.deployment.id, out.prediction_id, labels=["Good"])  # ground truth
client.deployments.add_data(dep.deployment.id, "labelled_this_week.csv")            # more labelled rows

for v in client.deployments.versions(dep.deployment.id):  # retrains appear as candidates
    print(v.version, v.status, v.trigger.reason, v.metrics.comparison)
client.deployments.promote(dep.deployment.id, version_id)  # or .reject(...), .rollback(...)

3. CLI

agentml run "Classify banana quality, accuracy >= 0.98" -f banana_quality.csv --follow
agentml list
agentml code <run-id> -o pipeline.py
agentml ask <run-id> "Which features matter most?"
agentml drift <run-id> new_data.csv --wait

agentml deploy <run-id> --wait
agentml predict <deployment-id> -f rows.csv --explain
agentml feedback <deployment-id> <prediction-id> Good,Bad,Good
agentml add-data <deployment-id> labelled.csv
agentml versions <deployment-id>
agentml promote <deployment-id> <version-id>
agentml rollback <deployment-id>

Errors and limits

All of these inherit from AutoMLError. Requests that only read data are retried automatically on network errors and 502/503/504.

Configuration:

• AUTOML_AGENT_API_KEY: the API key.
• AUTOML_AGENT_BASE_URL: the API root, default https://agentml.kitasid.dev/api.
• AUTOML_AGENT_CONFIG_DIR: where agentml login saves the key.

The pipelines, agents and models run on the AutoML-Agent service. This package is only the client.