Docs / For agents

Set a ceiling. Agents never spend past it.

One key in front of image, video, voice and text, on prepaid credit, with the spend limit enforced by the API rather than by the agent.

The ceiling

A limit is a parameter, not an instruction

The agent says what it wants and the most it may spend. Corent decides the rest.

tool: create{  "intent": "a 10-second vertical sunrise clip",  "max_cost_cents": 50} → video-air · 9:16 · 10s · 44¢ billed

Ceiling 50¢

Billed 44¢

6¢ unspent

One create call: 44¢ billed against a 50¢ ceiling, checked before it spent

Checked before it spends

The ceiling is enforced server-side, so a looping agent cannot argue its way past it.

No model, no tier

The agent describes the outcome. The router picks the model and sets the parameters.

The charge comes back

meta.cost_cents rides on every successful response, and every request keeps a receipt.

MCP

One paste connects an agent

Hosted, or run the package locally. Either way the key stays server-side.

Hosted, nothing to install

https://mcp.corent.tech/mcp Streamable HTTP

Any MCP client, config file

"mcpServers": {  "corent": {    "command": "npx",    "args": ["-y", "corent-mcp"],    "env": { "CORENT_API_KEY": "co_live_..." }  }}

Tool	What it does	What it spends
plan	Previews what Corent would do for an intent, and what it would cost.	Nothing
create	Takes an intent and a max_cost_cents ceiling, decides everything, generates.	Up to the ceiling
list_models	The models an agent can pin by name, with live status and quality score.	Nothing

HTTP

For agents that do not speak MCP

One request, one result, the cost on the response.

The request

curl -X POST \  https://api.corent.tech/v1/images/generate \  -H "Authorization: Bearer co_live_..." \  -H "Content-Type: application/json" \  -d '{    "prompt": "a matte black watch on marble",    "tier": "air"  }'

The response

{  "status": "completed",  "images": [{ "url": "https://..." }],  "meta": {    "model": "image-air",    "cost_cents": 2,    "duration_ms": 1500  }}

Deciding what to generate on its own? Send the raw request to /v1/intent/execute with confirm_estimated_cost_cents_up_to, and Corent picks the tier conservatively rather than letting an agent reach for the most expensive one.

Reach

Four kinds of output, one balance

Same key, same credits, the exact charge returned every time.

Output	Endpoint	What comes back	The charge
Image	POST /v1/images/generate	The image inline, usually. A slow render returns a job id instead.	meta.cost_cents, from 3¢
Video	POST /v1/videos/generate	202 and a job id. Read GET /v1/jobs/{id}/stream instead of polling.	On the job, from 18¢ a clip
Voice	POST /v1/audio/speech	The finished audio file, synchronously.	meta.cost_cents, 9¢ per 1,000 characters
Text	POST /v1/chat/completions	An OpenAI-compatible completion. Streaming and tool calls pass through.	corent.cost_cents, from 3¢ a response

Budget

What stops a runaway loop

Six limits, every one of them outside the agent.

Limit	What it does
Per-request ceiling	max_cost_cents on create, confirm_estimated_cost_cents_up_to over HTTP. Past it, Corent returns the plan and the estimate and spends nothing.
Per-key daily cap	New keys start at $10 a day, with a rate limit and a concurrent-job limit beside it. Raising a cap needs a dashboard login, so a leaked key cannot raise its own.
Prepaid balance	GET /v1/account/balance, in cents. Top up from $5 by card, $10 in crypto. No subscription.
Balance at zero	The request returns 402 and is not charged. Nothing generates until funds land.
Failed generations	Never billed, failover included. The receipt reads billed: false, cost_cents: 0.
Blind retries	Send an Idempotency-Key. A repeated key returns the original job instead of charging twice.

Two minutes in. Take a key.

Get a key Full API reference

Prepaid credits from $5. No subscription. Failed generations are never billed.