# Marcel Routing

Marcel Routing is a prepaid, OpenAI-compatible multi-provider AI gateway. Developers use one Marcel API key instead of supplying their own upstream provider keys.

API base URL: https://marcel-agent.com/api/v1

## Providers and capabilities

The service routes supported requests to OpenAI, Anthropic, Google Gemini, xAI, DeepSeek, and Perplexity. Availability varies by model and operation; consult the live public catalog instead of assuming every provider supports every capability.

- [Live public catalog, no authentication required](/api/catalog)
- [Complete router API reference](/docs)
- [Text API reference](/docs.md)
- [OpenAPI contract with request and response schemas](/openapi.yaml)

Capabilities include chat completions, custom function tools where supported, image generation, asynchronous video generation, realtime voice sessions, embeddings, moderation, reranking, text and document translation, and integrated search/data tools. The API reference describes the operation-specific limits.

## Authentication and usage

Operational requests require Authorization: Bearer YOUR_MARCEL_API_KEY. Never include the key in a public URL, browser code, shared log, or prompt. Keep it in the calling application's secret store.

Customers manage keys and usage in the signed-in portal. Billing uses prepaid credit and automatic reload. Eligible customers may receive $5 in trial credit once per permanent card fingerprint after card setup; automatic reload starts at a $2 balance. Check the signed-in billing page for current packs and eligibility.

## Model selection and tool conversations

Use a model ID from the live catalog. Preserve the assistant message with its tool_calls and return a tool message for every tool_call_id before continuing a conversation. Marcel formats tool calls; the client executes the tools.

Chat requests can include fallback_model naming a main model. Marcel makes at most one fallback attempt on transient upstream errors, rate limits, or timeout. Invalid requests, insufficient credit, and disconnected clients do not trigger fallback. The response model and billing refer to the successful model.

GPT-5.6 Sol uses reasoning_effort: none for function tools on Chat Completions. Chat SSE is currently buffered: Marcel sends keepalive comments while waiting, then emits content and indexed tool-call chunks. It is not upstream token-by-token streaming. Each provider attempt has a 90-second deadline; allow over 180 seconds in a client using fallback.

## Public information versus private operations

The home page, routing overview, documentation, text guides, OpenAPI contract, and public catalog can be read without JavaScript or authentication. Account information, API keys, balances, usage history, administration, and paid API calls remain protected.