- checks the request and the reply against your guardrails;
- checks the key, the models it may use, and its budgets and rate limits;
- sends it to the right provider, with fallbacks if that provider fails;
- adds your MCP tools, saved prompts and cache;
- records what happened, and what it cost, in the request log.
Send your first request
A provider credential, a key and one request, in five minutes.
API reference
Every endpoint, header and error.
Start before you register anything
You can use the AI Gateway as soon as your organization exists. Add a provider credential, create a key, and send requests; your organization guardrails govern that traffic. When you later register the AI system it belongs to and it has been approved and gone live, attach the key, and that AI system’s own guardrails take over. Attaching before it is live stops the key’s traffic until it is: see Virtual keys.What it supports
- Request formats: OpenAI chat completions and Responses, Anthropic Messages, and Gemini. See the API reference.
- Model providers: OpenAI, Anthropic, Bedrock, Google Gemini, Google Vertex AI and Azure OpenAI, each with your own credentials. See Provider credentials.
- Clients: the OpenAI, Anthropic and Google SDKs, Claude Code, Codex CLI, Gemini CLI, Cursor, and any MCP client. See SDKs and Coding agents.
- Streaming, in every format.
Where you manage it
- AI Gateway in the sidebar holds keys, provider credentials, the organization’s guardrails, MCP servers, prompts, the request log and settings. Its Overview has the address and code to copy, and the gateway’s health.
- Each AI system’s AI Gateway tab holds that AI system’s keys, limits, MCP tools, cache and content logging, and its requests.
- Reporting shows what the traffic costs and what the guardrails caught. See Reporting.