How-to · luxin.sh

Image generation your agent runs end to end.

Autonomous agents need media they can create, recover, and pay for without a human wiring providers first. Luxin is a hosted tool for exactly that: one command, no provider key, and every result comes back as a durable URL with a job the agent can look up and a receipt for what it cost.

Hand this to your agent, or point it at luxin.sh/skill.md to read the contract.

Full disclosure: this is Luxin's own guide to its product. We kept it to what is true and testable, including when a built-in tool is enough.

The loop an agent can run

Choose, create, recover, and pay.

The whole media loop an agent needs to run alone, from picking a model to accounting for the spend, behind one command and one response shape.

01

Choose and create in one call

The agent describes what it needs and Luxin picks a fitting model from more than 200 and can write the prompt. One call covers model choice, prompting, and generation, so the agent does not carry a provider matrix.

02

Get back media it owns

Every result is a durable URL on media.luxin.sh with a recoverable job id, a cost receipt, and one stable JSON shape. It stays valid after the run ends, so a later step or a different agent can reuse it.

03

Pay and recover within a cap

The agent can fund its own runs within a budget you set, and can recover a job by id after a crash or a new session. Spend is prepaid and model-priced, so it cannot run past the balance.

Real output

Durable work, not a chat artifact.

An agent that loses its output on the next turn cannot build on it. Every image here has a URL, a job, and a receipt that outlive the session. Generated by an agent through the public CLI.

Isometric packaging mockup, matte teal box on raw concrete, crisp studio lighting, minimal
MOCKUP
fal.bytedance-seedream-v4-text-to-image · $0.05
Editorial portrait, soft window light, 85mm lens, shallow depth of field, magazine cover quality
PORTRAIT
xai.grok-imagine-image-quality · $0.12
When to reach for it

When a built-in tool stops being enough.

reach for luxin when

Your agent produces media as part of a task.

The output has to be inspectable, recoverable, and payable, and you want one contract across images, editing, video, audio, and 3D. It runs the same whether the agent lives in Claude Code, Codex, CI, or your own loop, and the agent manages its spend within a cap.

a built-in tool is enough when

It is one throwaway image inside one chat.

A single image you will never recover, cite, or pay for programmatically is fine on a runtime's built-in generator. Luxin earns its place once the output has to survive the run and the agent has to handle its own budget.

Questions

What agent builders ask first.

What makes image generation “for agents” different?

An agent needs the result as durable work, not a preview that disappears on the next turn. Luxin returns a hosted URL, a recoverable job, a cost receipt, and one stable JSON shape, so the agent can look the work up later, reuse it, and account for what it spent.


Can the agent recover a job after a crash or a new session?

Yes. Every create returns a job id, and the agent can look that job up later for its status, assets, and cost, even from a fresh session with no prior state. Nothing is lost when the conversation ends.


Can the agent pay for its own work?

Yes, within a budget you set. The agent can fund its runs on its own, and a human can also top up with one Stripe click. Credits are prepaid and model-priced, so the agent cannot spend past the balance or your cap.


Does it work across different agent runtimes?

Yes. Luxin is a plain CLI and HTTP API with no runtime lock-in, so the same command runs from Claude Code, Codex, Cursor, a shell script, CI, or a custom agent. Point any agent at luxin.sh/skill.md and it can read the contract and start.


How does the agent choose a model?

It does not have to. Luxin auto-selects a model that fits the request from more than 200 across providers, and can write the prompt. If the agent needs a specific model or a provider-native control, it can name it, and rare controls stay available as validated parameters.


Is it just a wrapper around one provider?

No. It spans more than 200 models across multiple providers behind one contract, and covers image, editing, video, audio, and image-to-3D. The agent learns one tool instead of one integration per provider and per modality.


What does the agent get back from a create call?

A durable media URL, a job id, the credits and cost it used, and stable JSON with a trace id. The shape does not change between models, so the agent parses one response and moves on.

Start now

Give your agent a media tool it can run alone.

Read skill.md

Want your agent to always have it? Install the skill: npx skills add https://github.com/danielgwilson/luxin --skill luxin