API reference

Memtro exposes the OpenAI chat-completions shape at https://platform.memtro.com/v1, so any OpenAI-compatible client works unchanged. Authenticate with Authorization: Bearer sk-memtro-….

Endpoints

Method and path	Purpose
`POST /v1/chat/completions`	Chat, streaming or not, with tools. The main endpoint.
`GET /v1/models`	The model ids available to you: `memtro`, `auto`, `auto/<provider>`, and `provider/model` for every provider you have a key for.
`POST /v1/embeddings`	Embeddings through your OpenAI or Google key.
`POST /api/mcp`	Memtro as an MCP server.

Model ids

Id	Meaning
`auto`	Complexity routing across the configured ladder. See Model routing.
`auto/anthropic`, `auto/openai`, `auto/google`, `auto/claude-code`	Routing within one provider.
`memtro`	The organisation's default model (set under Providers). `brain` is accepted as an older alias.
`anthropic/claude-opus-5`, `openai/gpt-5.4`, `google/gemini-3.8-flash`	A specific model through your key for that provider. Bare ids such as `claude-sonnet-5` are inferred.
`claude-code/opus`, `claude-code/sonnet`, `claude-code/haiku`, `claude-code/fable`	Run through the Claude Code CLI with your Claude Code token or Anthropic key.
`compat:<label>/<model>`	An OpenAI-compatible endpoint you configured under that label.
Any id with `:agent` or `:chat`	Forces agent mode or single-reply mode for this request, e.g. `auto:agent`.

The `memtro` request field

Add a memtro object to the request body to control how Memtro behaves. Every field is optional. (brain is still accepted as the field name for older clients.)

{
  "model": "auto",
  "messages": [{"role": "user", "content": "Summarise open VIP tickets"}],
  "memtro": {
    "mode": "agent",
    "memory": "org",
    "retrieve": true,
    "tools": true,
    "show_progress": true,
    "engine": "native",
    "cache": true
  }
}

Field	Values	Default	What it does
`mode`	`chat`, `agent`	`chat`, or the organisation's default	`agent` runs the multi-step harness: plan, query every relevant system, cross-reference, answer once. See Agent mode.
`memory`	`org`, `personal`, `off`	`org` when you are in an organisation	Where facts learned from this exchange are stored. Acts as a ceiling: within an `org` request, facts about you personally are still classified as personal.
`retrieve`	boolean	`true`	Inject relevant memories and graph context into the request.
`tools`	boolean	`true`	Expose Memtro's memory tools and your connector tools to the model.
`show_progress`	boolean	`true`	In agent mode, stream tool activity as `reasoning_content` deltas.
`engine`	`native`, `opencode`, `claude-code`	the organisation's setting	Which harness runs agent mode.
`cache`	boolean	`true`	Serve identical connector calls from the short-lived result cache.

What comes back

Responses are standard chat completions plus a memtro field:

{
  "memtro": {
    "mode": "agent",
    "engine": "native",
    "routing": {"requested": "auto", "model": "anthropic/claude-opus-5", "complexity": 3, "method": "heuristic", "reason": "…"},
    "tools": [{"tool": "helpscout_get_conversation", "input": {"id": 2450}, "ok": true, "summary": "…"}]
  }
}

When streaming, the first chunk carries memtro with the mode, engine and routing decision; tool activity arrives as reasoning_content deltas while the model works, and a final chunk lists the tools used. Open WebUI, Cherry Studio and most chat UIs render the reasoning as a collapsible "thinking" block.

Tools

Server tools run inside Memtro: memtro_search, memtro_remember, memtro_explore, the job tools, memtro_delegate in agent mode, and one set per connector you have access to (hubspot_search, helpscout_get_conversation, postgres_query, and so on). You do not declare them; they are always there unless tools: false.
Client tools declared in the request's tools array work the OpenAI way: the response returns tool_calls for you to execute. Client tools are not available with claude-code/* models, and requests that declare them fall back to the native harness.

Streaming

Set stream: true for server-sent events in the OpenAI chunk format. Add stream_options: {"include_usage": true} for a final usage chunk. The API also accepts temperature, top_p, max_tokens or max_completion_tokens, stop and image content parts.

Errors

Errors use the OpenAI error envelope ({"error": {"message", "type", "code"}}). Common codes: invalid_api_key, provider_not_configured (add a key under Providers), model_not_found, routing_unavailable, unsupported_tools (client tools with a claude-code/* model), engine_unavailable.

Limits

The API is rate limited per client address and long requests (agent runs) may take minutes; keep client timeouts generous. Each agent request has a step budget (default 30) set under Dashboard → Routing.