Gateway API

Authentication

All requests must include your API key as a Bearer token in the Authorization header. You can find your key on the Profile page after signing in.

curl https://getmegabrain.com/api/gateway/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-4o-mini", "messages": [{"role": "user", "content": "Hello"}]}'

Chat Completions

Fully compatible with the OpenAI Chat Completions API. Both streaming and non-streaming modes are supported.

Example — non-streaming

curl https://getmegabrain.com/api/gateway/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-3-5-sonnet",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "What is the capital of France?"}
    ]
  }'

Example — streaming

curl https://getmegabrain.com/api/gateway/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-3-5-sonnet",
    "stream": true,
    "messages": [{"role": "user", "content": "Tell me a joke"}]
  }'

Example — OpenAI SDK

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: "YOUR_API_KEY",
  baseURL: "https://getmegabrain.com/api/gateway",
});

const response = await client.chat.completions.create({
  model: "anthropic/claude-3-5-sonnet",
  messages: [{ role: "user", content: "Hello!" }],
});

console.log(response.choices[0].message.content);

Auto Model

Set model to one of the Auto Model IDs to let the Gateway choose the best model for each request.

Response

The OpenAI Chat Completions shape, plus provider (the serving upstream), native_finish_reason, and a usage object with cost (USD actually charged), is_byok and cached / reasoning token details. Every reply carries an X-Generation-Id header you can look up at GET /v1/generation?id=. Streams end with exactly one usage chunk and data: [DONE].

Parameters

Every field of the Chat Completions request the gateway accepts, and what it does with it. “Forwarded” fields reach the model unchanged — whether the model honours them is what its supported_parameters in GET /v1/models says.

* required

Embeddings

Generate vector embeddings from text. Compatible with the OpenAI Embeddings API.

curl https://getmegabrain.com/api/gateway/embeddings \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "text-embedding-3-small",
    "input": "The quick brown fox"
  }'

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: "YOUR_API_KEY",
  baseURL: "https://getmegabrain.com/api/gateway",
});

const result = await client.embeddings.create({
  model: "text-embedding-3-small",
  input: "The quick brown fox",
});

console.log(result.data[0].embedding);

Audio

Speech to text and text to speech, in OpenAI's shapes, so the OpenAI SDK's audio.transcriptions and audio.speech work against the same base URL. Audio models are their own set, not catalogue models — GET /v1/audio/models lists them. An upload is at most 4 MiB, and both routes need a gateway key: there is no anonymous tier for audio.

curl https://getmegabrain.com/api/gateway/audio/transcriptions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F file=@meeting.mp3 \
  -F model=openai/whisper-1

curl https://getmegabrain.com/api/gateway/audio/speech \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -o speech.mp3 \
  -d '{
    "model": "openai/gpt-4o-mini-tts",
    "input": "One API, every model.",
    "voice": "onyx"
  }'

List Models

Retrieve the list of models available through the Gateway. The response follows the OpenAI models format.

curl https://getmegabrain.com/api/gateway/models \
  -H "Authorization: Bearer YOUR_API_KEY"

{
  "data": [
    {
      "id": "anthropic/claude-fable-5.1",
      "canonical_slug": "anthropic/claude-fable-5.1",
      "name": "Anthropic: Claude Fable 5.1",
      "context_length": 1000000,
      "architecture": { "modality": "text+image->text", "input_modalities": ["text", "image"], ... },
      "pricing": { "prompt": "0.000005", "completion": "0.000025", ... },
      "supported_parameters": ["max_tokens", "reasoning", "tools", ...],
      "reasoning": { "supported_efforts": ["max", "high", "low"], ... }
    },
    ...
  ],
  "total_count": 445,
  "links": { "next": "/api/gateway/v1/models?offset=100&limit=100" }
}

Filter and page with query parameters: offset, limit (≤ 1000), q, supported_parameters, input_modalities, output_modalities, providers, arch, context, min_price / max_price (USD per million prompt tokens), min_output_price / max_output_price, zdr, max_age_days, min_bench (0–100 Terminal-Bench), sort (newest, pricing-low-to-high, pricing-high-to-low, context-high-to-low, bench-high-to-low). total_count is the size of the filtered set before paging. The public models page offers the same filters as facets and mirrors them into its query string.

Keys, credits, generations

Everything else the gateway serves on the same base URL. Named keys (mb-v1-…) carry their own spend limit and reset window; a management key can create keys and read credits but cannot call inference.

curl -X POST https://getmegabrain.com/api/gateway/v1/keys \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"name": "ci-runner", "limit": 25, "limit_reset": "monthly"}'

Errors

Every error, on every endpoint, uses the OpenRouter envelope: error.code equals the HTTP status, error.message is human-readable, error.metadata.error_type is the canonical class (invalid_request, authentication, payment_required, permission_denied, not_found, rate_limit_exceeded, provider_unavailable, …) and error.metadata.megabrain_error_type the precise reason below. Upstream failures also carry provider_code and raw. A failure after a stream has started arrives as a final chunk with an error object and finish_reason: "error".

{
  "error": {
    "code": 402,
    "message": "API key has reached its spend limit",
    "metadata": {
      "error_type": "payment_required",
      "megabrain_error_type": "insufficient_credits",
      "reason": "api_key_over_limit"
    }
  },
  "error_type": "insufficient_credits"
}

OpenAPI

The whole contract — every endpoint, the request and response schemas, the parameter and error references above — is published as an OpenAPI 3.1 document, generated from the same source as this page. Point a client generator or an API explorer at it.

MCP server

The catalogue and this reference are also served over the Model Context Protocol, so an agent can look a model up or read a parameter's range without a browser and without anyone pasting documentation into a prompt. Four read-only tools: search_models, get_model, list_providers and gateway_reference. No key: it answers from the public catalogue and never sees your usage or spend.