trainingR3 · 5/5 gates · v4 live

Compile your agent into one fast model.

Trace your agentic system once. train.cloud distills its tools, prompts, and routes into a single trained model behind a drop-in OpenAI-compatible endpoint — then keeps improving it on a gated flywheel.

Start a training run Read the endpoint docs

▸ no infra to manage▸ swap one base URL▸ promote only past the gates

three commands

From a sprawling agent to a single model — in one pipeline.

No orchestration graphs to babysit at inference time. Compile the behavior, train it, and serve it as one model.

01 / compile

$ train compile ./agent

Trace the system

Point it at your running agent. It captures the tools, prompts, and decision routes into a portable training spec — your code stays your code.

02 / train

$ train run --rounds 3

Run the rounds

Full supervised training across R1→R3 with a live cost meter and a hard budget cap. Each round reports loss and a fresh eval scorecard.

03 / serve

$ train serve --promote v4

Ship the endpoint

Promote a version that clears every gate. You get a versioned, OpenAI-compatible endpoint and an API key. Roll back instantly if needed.

0.11

final eval loss · run #128

5/5

promotion gates cleared

1×

base URL to swap

flywheel round, gated

what you get

Built like a tool, not a dashboard.

Everything is addressable, versioned, and keyboard-reachable. The product speaks one vocabulary: workspaces, projects, runs, models, endpoints, keys, credits.

// the flywheel

It gets better while you sleep — but only past the gates.

Production traffic feeds the next training round. Nothing reaches your endpoint until it beats the live model on every gate. Improvement is automatic; promotion is earned.

serve→collect→train R+1→gate→promote↺

// gates

Five gates, zero surprises.

quality ≥ baseline

no regressions

latency budget

safety checks

cost per call

// endpoint

Drop-in compatible.

Change one base URL. Same request shape, same streaming, same SDKs. Chat, completions, and embeddings on a single versioned endpoint.

// keys

Keys you can actually manage.

Scoped API keys per project, instant rotation, last-used timestamps, and per-key spend. Reveal once, then it's hashed — like every key should be.

// credits

Spend you can predict.

One credit balance covers compile, training, serving, and storage. Set a hard cap per run; watch the burn-down stream in real time.

the endpoint

If you can call an LLM, you can call your model.

No new SDK to learn. Point your existing client at api.train.cloud, pass your key, and you're talking to the model you compiled — versioned, gated, and yours.

Python

from openai import OpenAI

client = OpenAI(
    base_url="https://api.train.cloud/v1",
    api_key="sk-traincloud-••••••••"
)

response = client.chat.completions.create(
    model="your-model-v4",
    messages=[{"role": "user", "content": "..."}]
)

Ready to compile your first agent?

Start a training run