Agents CLI

Every stateset-agents command, grouped by what it does — and which ones spend money.

One entry point, stateset-agents, installed with the package. Commands below are grouped by the part of the loop they belong to.

stateset-agents --help
stateset-agents version

The improvement loop

Command	Does
ingest	Convert OpenAI or LangChain logs into per-conversation transcripts
improve run	Grade, curate above a threshold, write curated.jsonl and next_steps.md
improve status	Read back a previous run’s summary
evaluate	Score a model against a reward and a prompt set
benchmark	Run the packaged benchmark suites

Training

Command	Does	Spends
train	Train locally against a pinned backend	Your own GPU
train-remote	Rent a GPU, train, evaluate, release it	Yes — bound by --max-cost
chat	Talk to a local model or adapter	No
chat-remote	Talk to a remotely hosted adapter	Yes, while the pod is up
undeploy	Tear down a deployment	No

train-remote and chat-remote rent hardware. train-remote refuses to start a run that could exceed --max-cost and terminates the pod on every exit path, including your machine dying mid-run. chat-remote holds a pod open for as long as the session lasts — close it when you are done rather than leaving a chat window open overnight.

Setup and diagnosis

Command	Does
init	Scaffold a project
doctor	Check the environment for the things that break training
validate-config	Check a config without running anything
preflight	The checks train runs before it starts, on their own
publish-check	Verify a build is publishable
provider-canary	Prove a remote provider is reachable and behaving

Run doctor before your first train, not after it fails. It catches the environment problems — driver, toolkit, disk, backend version — that otherwise surface an hour into a run as something that reads like a model problem.

Model presets

Named presets resolve a model id with its hyperparameters, so a training command does not carry twelve flags:

stateset-agents qwen3-8-27b --help

Presets ship for the Qwen 3 family, Kimi K2.6 and K3, Muse Glimmer, Nemotron 3.5, Qwen 3 Coder, GPT-OSS and DeepSeek V4. list_model_presets in the MCP server returns the same list with each preset’s key fields.

Serving and integration

Command	Does
serve	Run the FastAPI gateway — the Agents API
mcp	Run the MCP server
flywheel	Drive the continuous improvement loop
advanced	Lower-level entry points for the RL algorithms
research	The auto-research workflows

Rented GPUs

Four commands cover the whole rented-hardware lifecycle. Every pod is terminated on every exit path — success, failure, timeout, or your laptop dying mid-run — and every pod records what it cost.

Command	What it does
train-remote	Rents a GPU, fine-tunes, returns the adapter plus eval_results.json, gives the hardware back
chat-remote	A multi-turn conversation with a tuned model; every transcript is saved in the format ingest accepts, so a chat becomes training data
serve-remote	A persistent vLLM endpoint behind a bearer token, with --max-hours self-destruct and --list / --stop
costs	What every remote run actually cost — model, hardware, duration, dollars

Run costs before you believe a benchmark. A number that arrived without its bill is a number you cannot budget for, which is why the framework refuses to lose track of a pod.

Next steps

Agents quickstart

These commands in the order you would actually run them.

Overview

The loop, the rewards and the backend adapters.

MCP server

Driving the loop from an agent.

RL platform

Where this sits in the wider StateSet RL story.