API Reference

Create a chat completion. Supports streaming via SSE, tool calls, and the MCP agentic loop (set x-argus-mcp: true). OpenAI-compatible request and response shape.

Generate vector embeddings for input text. Routes to the configured embedding provider.

Generate images via the configured image provider (e.g. gpt-image-1).

Transcribe audio (whisper, gpt-4o-transcribe, ElevenLabs, Google Chirp). Multipart upload.

Synthesize speech (tts-1, gpt-4o-mini-tts, ElevenLabs).

List models THIS virtual key can call. Filter with q, provider (repeatable), capability (chat | embeddings | images | audio), limit, cursor. Response is the intersection of the gateway catalog and the key's allowedModels / providerMap.

Single-model detail — pricing, context window, capabilities, available_to_key. Returns 404 model_not_found when the key cannot reach the model.

Pricing table across all providers. Filter with provider (repeatable) and q. Prices normalised to per-1M tokens for token-billed models.

Key-scoped usage aggregates. Totals + time-bucketed series with start, end, bucket (hour | day | week), group_by (none | model | provider). Never returns cross-key data.

JSON-RPC 2.0 MCP gateway. Handles tool discovery, invocation, and circuit breaking across every registered MCP server. Supports initialize, tools/list, tools/call, resources/list, resources/read, resources/templates/list, resources/subscribe, prompts/list, prompts/get, sampling/createMessage and logging/*. Batch arrays up to 50 entries. Requires a virtual key of type mcp or all. POST /mcp is a legacy alias — use /v1/mcp.

Discoverable MCP catalog for the calling key's organization. Returns active servers with per-server tool_count.

Flattened tool list across visible MCP servers. Filter with server (name or id, repeatable), q, limit, cursor.

Invoke a specific MCP tool over REST instead of JSON-RPC. Body takes server (name or id), tool, and arguments — all three in the body, not the path.