# Prism Inference: Full Agent Context

Prism is a paid inference API for coding agents. It serves open-weight models
through OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages.

## Start here

- [Human and agent setup](https://docs.prisminference.com/guides/coding-agents)
- [Agent Skill](https://prisminference.com/skills/prism-inference/SKILL.md)
- [Documentation index](https://docs.prisminference.com/llms.txt)
- [Quickstart](https://docs.prisminference.com/quickstart)
- [Authentication](https://docs.prisminference.com/authentication)
- [SDKs and clients](https://docs.prisminference.com/sdks)
- [Models](https://docs.prisminference.com/models)
- [Rate limits](https://docs.prisminference.com/rate-limits)
- [Errors](https://docs.prisminference.com/errors)

## API contract

- OpenAI base URL: https://api.prisminference.com/v1
- Anthropic base URL: https://api.prisminference.com
- Bearer credential: PRISM_API_KEY
- Chat Completions: POST /v1/chat/completions
- Responses: POST /v1/responses
- Messages: POST /v1/messages
- Count Messages tokens: POST /v1/messages/count_tokens
- List models: GET /v1/models
- Retrieve model: GET /v1/models/{model}

Callable models:

1. `deepseek-v4.1-flash`: text and base64 image input, 1M context.
2. `gemma-4-31b`: text and base64 image input, 32K context, served at BF16.

Always query `GET /v1/models`. Do not send requests to models absent from the
response.

Context limits: `deepseek-v4.1-flash` accepts up to 1,000,000 tokens of context;
`gemma-4-31b` accepts up to 32,768. Prism does not truncate silently. Large
requests return a `context_length_exceeded` error instead of a partial result.

## Errors, retries, and rate limits

Errors use an OpenAI-compatible envelope with a stable machine-readable `code`,
a `retryable` flag, a `fix` hint, and a `docs_url`:

```json
{
  "error": {
    "message": "API key rate limit exceeded.",
    "code": "rate_limit_exceeded",
    "retryable": true,
    "fix": "Wait for the Retry-After interval, then retry with bounded exponential backoff.",
    "docs_url": "https://docs.prisminference.com/errors"
  }
}
```

Retry `408`, `429`, `500`, `502`, `503`, and `504` when the request is safe to
repeat. Do not retry other `4xx` responses until the request or credential
changes. A `429` carries `Retry-After` in seconds; honor it, then use bounded
exponential backoff with jitter. See
https://docs.prisminference.com/rate-limits.

## Prompt-first setup

The coding agent creates the account, receives its restricted key immediately,
and stores it without exposing it. The human only supplies the emailed OTP;
they never see or re-enter the key.

Required sequence:

1. Read the matching agent guide.
2. Inspect existing config without exposing secrets.
3. Collect the owner email and username, then send `human_email` and `username`
   to `POST https://prisminference.com/api/agent-signups`.
4. Immediately pass the returned `api_key` into the harness-specific native
   secret mechanism. Never repeat it to the human or include it in the final
   response.
5. Ask the human only for the emailed OTP and call `POST
   /api/agent-signups/verify` with the key as Bearer authentication.
6. Show a redacted preview and make the smallest targeted provider change.
7. Validate the configured provider and model using the agent's native command
   and live model list.

Native key destinations:

- Codex: the `PRISM_API_KEY` environment selected by `env_key`.
- OpenCode: its auth store, written through `/connect` with custom provider ID
  `prism`.
- Hermes: `~/.hermes/.env`, written by `hermes config set PRISM_API_KEY`.
- OpenClaw: its auth store, written by piping the key to `openclaw models auth
  paste-api-key --provider prism` on stdin.

Agent guides:

- Codex: https://docs.prisminference.com/guides/codex
- Cursor: https://docs.prisminference.com/guides/cursor
- OpenCode: https://docs.prisminference.com/guides/opencode
- Hermes: https://docs.prisminference.com/guides/hermes
- OpenClaw: https://docs.prisminference.com/guides/openclaw

## Compatibility boundaries

Chat Completions supports streaming, function tools, structured outputs,
reasoning controls, and supported base64 images on DeepSeek V4.1 Flash.
Responses is stateless. Prism does not store responses, chain
`previous_response_id`, run background jobs, or provide hosted web search,
file search, computer, shell, local shell, or native apply-patch tools.

Anthropic Messages supports streaming, tools, thinking controls, and token
counting. Prism accepts Bearer or `x-api-key` authentication but Bearer is the
recommended common format.

## Billing and security

Prism is prepaid. Configuration and model-catalog validation do not consume
inference credit. Any live model request does. The human funds the workspace in
the browser. Agents must not access payment methods, enable auto-top-up, or
expose credentials.

Prism applies zero data retention to API inputs and outputs and does not use
them for model training.

- Privacy: https://prisminference.com/privacy
- Terms: https://prisminference.com/terms
- Billing: https://prisminference.com/app/settings/billing
- API keys: https://prisminference.com/app/settings/api-keys

## Rate limits

Per-key rate limits return `429` with a `Retry-After` header in seconds. Honor
`Retry-After`, then retry with bounded exponential backoff and jitter. See
https://docs.prisminference.com/rate-limits.

## Machine-readable

- [OpenAPI 3.1](https://docs.prisminference.com/openapi.yaml)
- [API catalog](https://prisminference.com/.well-known/api-catalog)
- [Security contact](https://prisminference.com/.well-known/security.txt)
- [Pricing](https://prisminference.com/pricing.json)
- [Status](https://status.prisminference.com)
- [Documentation index](https://docs.prisminference.com/llms.txt)