Skip to main content
Studio’s model catalog uses AWS Bedrock. The desktop does not call Anthropic’s hosted API directly and does not offer a per-user API-key path that bypasses the configured Bedrock route.

​
Current models

The current catalog includes: Model availability can change by Studio release and Bedrock region. The selector in Studio is authoritative for the installed build.

​
Inference regions

Choose the user-facing region under Settings → AI → Bedrock region: Changing region resets the current conversation’s prompt cache. The next call sends its full uncached input once.
The inference setting does not relocate Studio’s identity, billing, organization records, or control-plane infrastructure. Treat model-processing region and application-data residency as separate review items.

​
One Copilot turn

1

Assemble context locally

Studio combines the prompt, conversation history, organization and channel grounding, applicable memories, tool schemas, and the active objects you explicitly attached or mentioned.
2

Apply context safeguards

Studio removes known secret patterns and excludes normal credential payloads in favor of references. Redaction is defense in depth, not permission to paste a secret.
3

Build cacheable layers

Stable system and organization context precedes conversation and turn-specific content so Bedrock prompt caching can reuse eligible prefixes.
4

Sign and send

Studio signs the Bedrock request with scoped short-term AWS credentials and sends it over TLS to the resolved Bedrock endpoint.
5

Stream and gate tools

Bedrock streams the response. Proposed tool calls pass through effective policy and the approval gate before dispatch.
6

Return tool results

Tool results are added to the agentic loop as context for the next model call. A result from a host, connector, MCP server, browser, or generated app is untrusted input until validated.

​
What reaches model context

​
Local and cloud boundary

Local execution does not mean its result stays local. When you send terminal output, a screenshot, a transcript, or a file to Copilot, that selected content becomes cloud inference context.

​
Prompt caching and compaction

Prompt caching reduces repeated input processing when the prefix remains stable. Region, model, system context, and some policy changes can invalidate the cache. Compaction replaces older conversation detail with a summary when a chat grows. Run /compact [focus instructions] when you want to preserve a named objective while dropping low-value history. For models with a selectable context window, 200K summarizes sooner and 1M retains more at materially higher long-session cost. No summary is lossless. Start a new chat with explicit artifacts when exact historical detail matters more than continuity.

​
Browser AI Grounding

Browser AI Grounding adds visual browser context to help Copilot interpret complex pages. It consumes additional tokens per action and can include visible page content. Enable it only for the browser session and task that need it.

​
Thinking and effort

Supported models expose adaptive thinking and effort controls. Thinking state can be visible while a run executes; the durable transcript focuses on messages, tools, approvals, and results. Do not rely on an internal reasoning trace as an audit control—rely on the displayed action and recorded external evidence.

​
Data-handling guidance

  • Attach the minimum context necessary for the task.
  • Use Key Chain references rather than plaintext secrets.
  • Treat connector, MCP, browser, website, and file content as untrusted input.
  • Verify model-generated conclusions against live read-only evidence.
  • Check the configured region and model before regulated work.
  • Review Known limits and roadmap as part of security approval.

Human in the loop

Understand response modes, tool classes, Autopilot, and other approval surfaces.

Agent and local runtime

See how the desktop, renderer, helper, and cloud services divide work.