AI Copilot: Providers, Context, and Reviewed Edits

Open hosted Studio at /editor. The website is / and documentation starts at /docs/vinyasa/overview/. No Vinyasa installation or local server is required.

Vinyasa supports three distinct AI workflows: built-in Copilot sends bounded inspection requests directly from your browser to a configured AI provider server; external chat assistants use the downloaded reference to return document YAML; MCP agents connect to the hosted document tools at /mcp. MCP does not supply an AI model and its public endpoint needs no provider key. All workflows share document validation and the element catalog. AI proposes changes; it does not certify a design or automatically change your canvas. Use synthetic content and review the data flow before sending a draft.

1. Open YAML and AI from the editor rail, then choose the AI Copilot tab.
2. Open AI configuration and select Google Gemini, OpenAI, Anthropic, or OpenAI-compatible / local. Enter the exact model name and credentials required by that endpoint.
3. Set Model context window (tokens) to the model/runtime limit. The default is 50,000, and the value is preserved when switching providers. Save configuration keeps settings in tab memory; it does not test a connection.
4. Choose Modify draft to edit the current YAML, or Replace document to start from a blank document. Replacement is still only a proposal.
5. Write one focused request, review the consent checkbox, and choose Generate proposal.
6. Copilot inspects schemas and relevant objects in bounded steps. It may make multiple provider calls; Cancel stops current and subsequent requests.
7. Review the YAML diff, then choose Apply proposal or Discard. Applying is explicit and undoable. If the canvas changed, reload before applying a stale draft.

Keys are not stored in localStorage and are not exported with projects. They are transmitted to the configured provider for authentication. Model names, availability, billing, and browser-access support depend on that provider. Start with the provider's official console: Google AI Studio, OpenAI platform, or Anthropic console. Local endpoints have additional protocol and browser requirements.

For ChatGPT, Claude, Copilot, or another external assistant, attach vinyasa-document-reference.md from Help > Download Syntax Ref. It contains complete document and .rvv envelope examples, page/settings fields, common properties, frame/group nesting, coordinate rules, validation limits, production URLs and generated property constraints. Ask for bare document YAML with schemaVersion, elements, pages and settings, explicit stable IDs, and no invented CSS fields. Paste the result into the YAML tab, validate, inspect it, and choose Apply to canvas. A bare document is not an importable .rvv project envelope.

For a tool-aware agent, connect Streamable HTTP to /mcp. The hosted server exposes list_elements, get_element_schema, validate_document, edit_document, and export_project. It requires no token or OAuth; it is unrelated to your AI provider key. Follow the MCP guide for VS Code/Claude Code configuration, exact tool arguments, response extraction, and troubleshooting. Tools process supplied documents; they cannot read browser projects or apply changes to your canvas. Your client controls approvals and may send results to its provider.

The built-in agent starts from counts, type names, and a bounded page overview rather than sending all YAML. It searches summaries, reads selected properties by ID, and asks for schemas as needed. Each turn retains a short plan and the latest inspection instead of an ever-growing transcript. Updates merge locally, preserving untouched properties.

Generation is capped at ten requests, with a 60-second timeout per request and 4,096 output tokens. Input is conservatively budgeted; oversized prompts or inspections fail explicitly. Unread-element edits and invalid properties are rejected, not applied or rerouted to another provider. Large text fields may require direct YAML editing. Read large-document limits before requesting a broad redesign.

For external prompts, ask for stable IDs, explicit left/top positions, supported properties, synthetic records, and validation before presentation. Both frame and user-group support children; windows and cards are visual elements, not implicit containers. Top-level coordinates are workspace-relative, while child coordinates are container-local. pageId is ownership metadata, not a position offset.

These are implementation presets, not promises of current availability or account access. Verify your provider's supported model ID and billing. Enter the base URL, not an already-completed request path, to avoid duplicated routes. Gemini uses x-goog-api-key; Anthropic uses x-api-key plus its version and direct-browser-access headers; OpenAI/compatible uses Bearer authorization when a key is supplied. Managed providers require a nonempty key; compatible endpoints follow their server's authentication rules. Never put keys in endpoint URLs, documents, screenshots, Git or MCP configuration.

Gemini must return completed JSON; OpenAI/compatible must support JSON-object output and a completed chat choice; Anthropic must return the forced vinyasa_action tool result. Arbitrary text or fenced YAML is not the built-in agent protocol. Use HTTPS provider servers; URLs containing embedded credentials, query strings or fragments are rejected. Requests run directly in the browser, so the provider must permit the hosted Vinyasa origin. There is no hosted proxy or automatic fallback. See compatible server requirements.

Cancel aborts active and subsequent requests but cannot recall already-sent data or provider charges. Errors and discarded proposals leave the applied canvas unchanged. Export an .rvv backup before major changes. Use MCP troubleshooting for external agent transport problems, not the provider settings above.

Try a focused request such as: "Find the billing screen's Save button. Keep its ID, text, size, and color; move it 16 pixels down. Do not modify other screens." For new layouts, specify screen size, controls, spacing, and the intended audience. Review the prompt recipes and login demo.

Vinyasa produces prototypes, not production applications. If the next step is shared design-system work or a code-backed application, consult the tool comparisons rather than assuming one workflow replaces all others.