Orkestrateorkestrate
Create your persona

Model & provider disclosure

Which models and providers Orkestrate uses to generate persona replies.

Last updated: June 13, 2026.

Inference

Orkestrate is model-agnostic. Replies to persona invocations are generated by a language model accessed through an OpenAI-compatible API endpoint.

  • Default provider: AWS Bedrock Mantle (us-east-1), reachable at https://bedrock-mantle.us-east-1.api.aws/v1.
  • Default model: mistral.ministral-3-3b-instruct, a small open-weight instruction-tuned model chosen for fast inference, predictable tone, and cost efficiency at MVP scale.
  • Configuration: the provider base URL is set via the OPENAI_BASE_URL environment variable, the API key via BEDROCK_MANTLE_API_KEY or OPENAI_API_KEY, and the model via BEDROCK_MANTLE_MODEL. Operators can swap provider or model without a code change.
  • Limits: requests are restricted to 600 completion tokens and a 25 s timeout. Sampling temperature is 0.7.

What is sent to inference

For each persona reply, the following inputs are transmitted to the inference provider:

  • The persona's soul document (Markdown text).
  • The user's current message, verbatim.
  • Up to 6 recent persona-invocation exchanges (user message + persona reply).
  • Static facts and dynamic context retrieved from long-term memory.
  • The current speaker (user or host) so the model can position the reply.

Not sent:raw persona traces (stored separately in Orkestrate's database), bearer tokens, GitHub identity, or memory summaries that aren't relevant to the current message.

Long-term memory

Persona continuity relies on Supermemory AI as a long-term memory store. After each invocation, a one-line summary (user: …\npersona: …) is written to a per-persona memory container (<userId>:<personaId>). On subsequent invocations, the memory store returns static facts, dynamic context, and a small set of relevant memories to inform the reply.

Memory is not shared between users, between personas belonging to different users, or between personas belonging to the same user unless explicitly scoped. Each persona has its own memory container.

Authentication provider

Identity is provided through GitHub OAuth, brokered by Supabase Auth. Orkestrate never sees your GitHub password. We receive your GitHub user id, username, display name, avatar URL, and primary email on sign-in.

Analytics & operational telemetry

Orkestrate does not ship third-party analytics to the browser. Server-side access logs and trace records are maintained for debugging and abuse mitigation. Traces can be deleted on request.

Changes

We may change provider or model as the service evolves. Material changes — for example, swapping providers — will be reflected on this page.

Contact

Questions about providers or models: hello@orkestrate.space.