API documentation

Rocket Relay is compatible with OpenAI-style client integrations. Point your SDK to the relay endpoint, authenticate with a Rocket Relay API key, and route traffic across models with one integration surface.

Quick start

Keep your existing client SDK and replace the base URL with Rocket Relay.

Standard balance billing

https://relay.rcapalpha.com/v1

BYOK routing (use credentials from your own pool)

https://relay.rcapalpha.com/byok/v1

Point any OpenAI-compatible client at /byok/v1. The upstream provider is inferred from the requested model, just like the shared pool — call a Claude model and your Anthropic key is used; call a GPT model and your OpenAI key is used.

CLM System One on Jetson

Use model clm-latest for choice scoring and ranking. It runs on our Jetson service through Rocket Relay and uses your normal Relay API key and balance. The price is $0.30 per 1 million input tokens; these operations produce no billable output tokens.

POST https://relay.rcapalpha.com/v1/rank accepts context, question, and an answers array.

curl -sS 'https://relay.rcapalpha.com/v1/rank' \
  -H 'Authorization: Bearer YOUR_RELAY_API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{"model":"clm-latest","context":"An invoice for cloud hosting","question":"Choose the category","answers":["billing","travel","support"]}'

POST https://relay.rcapalpha.com/v1/systemone accepts the official CLM state and questions format. Both endpoints are limited to 64 KB per request. API keys with explicit scopes need proxy:clm or proxy:*.

DeepSeek and Ollama

Both providers are supported through Rocket Relay's OpenAI-compatible provider bridge and use the same client endpoints as the rest of the gateway.

Provider	Upstream configuration	Call names	Catalog behavior
DeepSeek	Official API with an API key	deepseek-v4-flash deepseek-v4-pro	Listed in the shared-price catalog when the route and pricing are active.
Ollama	Reachable Ollama server; an API key is usually not required	Configured public alias or discovered upstream model ID	Deployment-specific. Query the runtime model endpoint for the exact list.

Connecting a compatible provider does not automatically make every discovered model a prepaid shared model. A model also needs an active pricing record before it appears in the public sell-side price catalog and can use balance billing.

Haiku 5.5 uses the short-context token tariff (prompts up to 100k tokens). Argon uses introductory rates through November 30, 2026. Long-context, cache-write and tool surcharges are not separately billed by the current token tariff.

Operators add these routes under Admin, CLIProxyAPI configuration, API / local model access. Saving a route hot-reloads the proxy configuration.

Authenticated runtime source of truth:GET https://relay.rcapalpha.com/v1/models

Code examples

curl https://relay.rcapalpha.com/v1/chat/completions \
  -H "Authorization: Bearer sk-rr-YOUR-KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-4-6",
    "messages": [{"role": "user", "content": "Summarize our changelog"}]
  }'

Reasoning and thinking controls

Rocket Relay normalizes reasoning controls across OpenAI, Claude, Gemini, Kimi, DeepSeek, and xAI-compatible upstreams. The relay recognizes the controls below, but the exact values accepted by a request depend on the selected model.

Control	Standard budget equivalent	Behavior
none	0	Disables reasoning when the selected model allows it.
auto	-1	Uses dynamic reasoning, or a model-appropriate fallback when dynamic mode is unavailable.
minimal	512	Minimum standardized reasoning effort.
low	1,024	Fast responses with lighter reasoning.
medium	8,192	Balanced speed and reasoning depth.
high	24,576	Greater reasoning depth for complex tasks.
xhigh	32,768	Extra-high reasoning depth on supported models.
max	128,000	Maximum available reasoning depth on supported models.

OpenAI-compatible request

curl https://relay.rcapalpha.com/v1/chat/completions \
  -H "Authorization: Bearer sk-rr-YOUR-KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-sol",
    "messages": [{"role": "user", "content": "Solve this carefully"}],
    "reasoning_effort": "high"
  }'

Responses API request

curl https://relay.rcapalpha.com/v1/responses \
  -H "Authorization: Bearer sk-rr-YOUR-KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-sol",
    "input": "Solve this carefully",
    "reasoning": {"effort": "high"}
  }'

You can also append a control or numeric token budget to the model name, for example gpt-5.6-sol(high), gemini-3.8-flash(minimal), claude-opus-4-8(max), or gemini-2.5-pro(8192).

Typical model-family support: OpenAI/Codex models generally accept low, medium, high, and xhigh, with selected newer models adding max; Claude adaptive-thinking models use low through max; Gemini 3 models use a model-specific subset of minimal through high; Kimi and most xAI reasoning models use low, medium, and high; DeepSeek V4 uses high and max.

Unsupported request-body levels return a validation error for the target model. Numeric budgets and model-name suffixes may be converted or clamped to that model's supported range. ultra is not currently accepted by the relay's unified reasoning parser; use max as the highest documented control.

Image generation and edits

Use GPT Image 2.5 Flare for image generation or Sunburst for precise edits. Check GET /v1/models for the call names available to your key.

Image call names: gpt-image-2.5-flare, gpt-image-2.5-sunburst, gpt-image-2.5, gpt-image-2.

Generate an image

curl https://relay.rcapalpha.com/v1/images/generations \
  -H "Authorization: Bearer sk-rr-YOUR-KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2.5-flare",
    "prompt": "A clean product render of a transparent glass teapot on a walnut table",
    "size": "1024x1024",
    "response_format": "b64_json"
  }'

Edit an image

curl https://relay.rcapalpha.com/v1/images/edits \
  -H "Authorization: Bearer sk-rr-YOUR-KEY" \
  -F model="gpt-image-2.5-sunburst" \
  -F image="@input.png" \
  -F prompt="Replace the background with a bright studio setup" \
  -F response_format="b64_json"

GPT Image 2.5 defaults are $4 input, $1 cached input, and $15 output per 1M reported tokens. Relay applies one input rate to text and image tokens; check the model catalog for your configured sell-side prices. The generic gpt-image-2.5 call name uses the same Relay family rate.

BYOK base URLs

If your tool only supports an OpenAI-style base URL and API key, switch the base URL to a BYOK endpoint instead of adding custom headers.

Use case	Base URL
Standard shared upstream billing	https://relay.rcapalpha.com/v1
BYOK (use credentials from your own pool)	https://relay.rcapalpha.com/byok/v1

`/byok/v1` is the single BYOK entry point. The upstream provider is chosen from the model name (e.g. `gpt-5.5` → your OpenAI key, `claude-sonnet-4-6` → your Anthropic key). Requests that map to a provider you have not stored a credential for return an error — they are never served from the shared account pool.

Endpoints

Method	Endpoint	Description
POST	/v1/chat/completions	OpenAI-compatible chat completions
POST	/v1/messages	Claude Messages API compatibility
POST	/v1/embeddings	Embeddings and vector-ready workflows
POST	/v1/images/generations	GPT Image 2 image generation
POST	/v1/images/edits	GPT Image 2 image edits with multipart or JSON image inputs
GET	/v1/models	Runtime models available to the authenticated API key
ANY	/byok/v1/...	Same OpenAI-compatible endpoints, but forced through your stored BYOK credential pool (provider chosen from the requested model)

Rate limits

Tier	RPM	TPM
Starter	60	100,000
Professional	300	500,000
Enterprise	1,000	2,000,000

Response headers include `X-RateLimit-Limit`, `X-RateLimit-Remaining`, and `X-RateLimit-Reset`.

Active shared models and sell-side pricing

This table is generated from active pricing records. It includes DeepSeek when enabled, but intentionally does not enumerate deployment-specific Ollama models. Use the runtime model endpoint above for exact availability.

Loading model catalog...