NoCam APIFor AI agents
Everything an AI agent needs to use this API without guessing. Hand this page to yours.
This page is for the model doing the work. If you're a person, paste the prompt below into your agent, and it will take it from there.
Hand it to your agent
Use the NoCam API to make photos, videos and text.
Read https://docs.nocamclub.com/llms.txt first, and follow the rules on https://docs.nocamclub.com/for-ai-agents.md
My key is in the NOCAM_API_KEY environment variable. Never print it.
Before spending more than 100 credits in one go, tell me the price and wait for my yes.What to read
| What | Address |
|---|---|
| The map: every page in one line | /llms.txt |
| Everything in one file | /llms-full.txt |
| Any page as Markdown | Add .md: /api/generations.md |
| The API as OpenAPI 3.1 | /openapi.json, also at https://api.nocamclub.com/v1/openapi.json |
| What's on sale right now, with options and prices | GET https://api.nocamclub.com/v1/models |
| A skill file for Claude Code and similar agents | /skills/nocam/SKILL.md |
The rules
- Don't guess a model's options. They differ from model to model, and they change with what's on sale. Read
GET /v1/models/{id}:input_schemais the truth, andpriceslists what can be bought right now. An unknown field is refused, not ignored. - Check before you spend.
POST /v1/estimatetakes the same body asPOST /v1/generations, checks everything and gives the price. Nothing runs and nothing is charged. - Always send an
Idempotency-Key. Use a new one for each thing you want made, and the same one when you retry it. Then a timeout can't make you pay twice. - Don't start something again to check on it. Starting is
POST, and it costs money. Checking isGET /v1/generations/{id}, and it's free. Keep theid. - Photos: use
Prefer: wait=60. Videos: poll. Photos usually finish within a minute. Videos take minutes, so poll everypoll_after_ms. - Read
error.codeand act on it. Every error, on a refused call and on a failed generation alike, hashint(what to do) andretryable(whether the same request can work: on a failed generation, send it with a newIdempotency-Key).docs_urlhas more, but you don't need it to decide. - Check
status, not just the HTTP code. A generation that fails while you wait withPrefer: waitstill answers200, withstatus: "failed". - Tell the person when something was refused. Don't reword a refused prompt to slip past the rules. Refusals count, and twenty in a day pause the account.
- Download what should last. Result links stop working at
expires_at, 7 days by default. Uploaded files go after 48 hours. - Never top up on your own. On
insufficient_credits, send the person totop_up_url. - Never print or commit the key. Read it from the environment.
The flow
GET /v1/models What's on sale?
GET /v1/models/{id} What does this model take, and what does it cost?
POST /v1/uploads Files: pass {url}, or get a link to PUT to
POST /v1/estimate Is this valid, and what's the price?
POST /v1/generations Start. Charged now. Send Idempotency-Key.
GET /v1/generations/{id} Until status is succeeded, failed or canceled.status goes queued → running → succeeded, failed or canceled. A generation that fails or is canceled gives its credits back by itself: don't try to claim them.
What to do with an error
In short, with the same meaning as retryable and hint:
| What you see | What it means | What to do |
|---|---|---|
invalid_parameter | A field is wrong. path names it and hint lists what works | Fix it and send again |
unknown_model, not_for_sale, no_route | Not available with these settings | Pick from prices in GET /v1/models/{id}, or another model |
input_refused, content_refused | Against the content rules, or the model refused | Tell the person. Change the idea, don't reword around it |
insufficient_credits, spend_limit_reached | Out of credits or over a limit | Tell the person. Send them to top_up_url |
rate_limited, too_many_active | Too fast or too many at once | Wait for Retry-After, then send the same request |
provider_unavailable, generation_failed | The model failed. Credits are back | Try once more with a new key, then another model |
timed_out | We stopped waiting. Credits are back | Check the same id later: the result may still come, free |
internal | Our fault | Retry with the same Idempotency-Key |
invalid_api_key, missing_scope, account_suspended | The key or the account | Stop and tell the person |
All of them: Errors.
Which model
Short version below. Details and prompt tips are on each model's page, starting at Models.
| The job | Use |
|---|---|
| A photo of someone or something from one photo, cheap and good | nano-banana-2 |
| The same character across many photos | nano-banana-pro |
| A poster or anything with exact words on it | gpt-image-2 |
| Many reference photos at once, cheap | seedream-5.0-flash |
| A photo from text alone, as cheap as it gets | grok-imagine |
| A talking clip from one photo, 10 seconds with sound, nearly free | muse-video |
| A short clip from a photo or text, or changing a clip you have | omni-flash |
| The best clip, up to 30 s, with photos, video and audio as references | seedance-2.5 |
| The same, cheaper, up to 15 s | seedance-2.0-fast |
| A text answer | gpt-6-luna, or claude-opus-5.5 for harder work |
Always confirm a model is on sale with GET /v1/models before relying on this table.
Writing prompts
- Name your files.
img1,img2,video1,audio1, in the order you sent them. See Files in prompts. - Say what to keep and what to change. Models tend to carry over anything you don't mention, like the background and the clothes in a photo.
- Put speech in quotes, and keep it short enough to say in the clip's length: about 2 words a second.
- Ask for simple movement. One action, a still camera. Complicated physics is where clips break.
- Spell out exact words. For text in a photo, write the words and add that the spelling must be exact.
Connecting through MCP
If your client speaks MCP, you don't need to write any HTTP. See MCP. A tool's error comes back as a result you can read, with the same codes and hints.
Testing without spending
A nc_test_… key runs the whole API on sample results at no cost, and the word fail in a prompt makes the generation fail so you can test your error handling. See Sandbox.