Inference paths
Use the installed model selector for current chat choices. Supported models include Claude and GPT families, with different context and effort options. Auto can route individual turns to different models. OpenAI models currently use a 200K effective conversation window; the extended-window option is only available for models that support it in Studio.
Studio does not require you to supply a personal provider API key for these managed routes.
Bedrock regions
Choose Settings → AI → Bedrock region for supported desktop Bedrock workflows:
Changing this setting resets the relevant conversation prompt cache. It does not relocate identity, billing, organization records, Org Brain indexes, cloud workers, or OpenAI processing. Review the processing boundary of each workflow separately.
One Copilot turn
1
Assemble context
Studio combines your prompt, conversation history, organization and channel grounding, selected memories, tool schemas, and attached or active objects relevant to the task.
2
Apply safeguards
Supported credentials are represented by references. Known secret patterns are filtered on supported paths. Redaction does not make it safe to paste a secret into a prompt, file, or tool result.
3
Send through the selected provider route
Bedrock requests use the configured AWS route; OpenAI requests use Studio’s authenticated proxy. Stable prompt layers can benefit from provider caching.
4
Stream and authorize tools
Proposed calls pass through the effective tool policy and applicable approval gate. Tool results become input to later model turns.
Org Brain data flow
Org Brain is a cloud retrieval system. Eligible source changes are sanitized, split into passages, embedded, and indexed. Source permissions remain part of the retrieval boundary; an organization-wide index is not permission to read every document in it.search_org_knowledge retrieves authorized evidence. ask_org adds a synthesized answer with citations when answer synthesis is enabled. The service can return insufficient evidence or refuse a request instead of producing an answer. A retrieved passage is untrusted content and cannot grant new tool permissions.
Answer-synthesis consent
The owner reviews the current policy under Administration → Organization. The supported policy disclosures are:
The configured policy determines which choice is offered. A policy or lifecycle change can require renewed consent. The desktop model picker and Bedrock region selector do not override the answer service’s disclosure. Do not interpret global OpenAI processing as a US-only residency commitment.
See Org Brain for activation, supported sources, citation inspection, dream cycles, and deletion behavior.
What reaches model context
Local work and cloud execution
Terminals, local shell, packet capture, and workstation Computer Use execute through the desktop and local helper. Their selected output can still be sent to cloud inference. Local semantic search and Org Brain are separate systems; the presence of local embeddings does not make Org Brain local. Cloud workers do not depend on an open desktop. They use the configured cloud tools and service credentials under their channel’s policy. A worker does not automatically gain access to an operator’s local shell, private credential, or desktop session.Caching and compaction
Caching can reduce repeated input processing. Model, region, prompt, and policy changes can invalidate cached prefixes. Compaction replaces older conversation detail with a summary as context grows;/compact [focus instructions] lets you identify what must be retained.
A summary is not an exact archive. Keep source artifacts and inspect citations when exact evidence matters. Do not use a model’s internal reasoning display as an audit record; use tool actions, approvals, transcripts, and the dedicated audit surfaces.
Review before rollout
- Confirm the selected provider and each enabled background workflow.
- Review Org Brain owner consent, source visibility, and answer-processing disclosure.
- Keep secrets in Key Chain and attach only the context needed for the task.
- Treat retrieved evidence and external tool results as untrusted input.
- Compare known limits with your retention, residency, and audit requirements.