AI Assistant

The in-product LLM assistant. Helps users write SQL, generate RLS policies, debug errors, explore the schema, and answer documentation questions. Integrates Model Context Protocol so it can call tools.

Studio surfaces involved

Folder	Purpose
apps/studio/components/interfaces/Docs/	The docked Assistant panel.
apps/studio/components/interfaces/SQLEditor/	Inline AI completions and "explain this query."
apps/studio/components/interfaces/Auth/Policies/	AI-generated RLS policy suggestions.
apps/studio/components/interfaces/Advisors/	LLM-summarized advisor findings.

Data layer

Under apps/studio/data/ai/. The Assistant uses Vercel's AI SDK (ai, @ai-sdk/openai, @ai-sdk/amazon-bedrock, @ai-sdk/react) plus the streaming functions in packages/ai-commands/edge.

MCP

Studio bundles @modelcontextprotocol/sdk and @supabase/mcp-server-supabase. This lets the Assistant call typed tools (e.g. "list tables", "run SQL") backed by Studio's existing data layer. @supabase/mcp-utils provides shared helpers for those tool implementations.

How a chat turn works

sequenceDiagram
    participant UI as Docs / Assistant panel
    participant Hook as data/ai/* hook
    participant Edge as ai-commands/edge
    participant LLM as OpenAI / Bedrock
    participant Tools as MCP tools (Studio data)

    UI->>Hook: user message
    Hook->>Edge: streaming command
    Edge->>LLM: prompt + context
    LLM-->>Edge: response stream (may include tool calls)
    Edge-->>Tools: dispatch tool calls
    Tools-->>Edge: tool results
    Edge-->>Hook: streamed tokens
    Hook-->>UI: render

Streaming runs on the edge runtime so first-token latency stays low; non-streaming flows use 'ai-commands' directly.

Evaluation

apps/studio/evals/ contains the Braintrust eval harness for the Assistant:

• pnpm --filter studio evals:run — run locally.
• pnpm --filter studio evals:upload — upload runs.
• pnpm --filter studio scorers:deploy — deploy online scorers.

CI workflows: braintrust-evals.yml, braintrust-preview-scorers-deploy.yml, braintrust-preview-scorers-cleanup.yml, braintrust-scorers-deploy.yml.

Integration points

• packages/ai-commands — prompt logic and streaming.
• AI SDK from Vercel for client-side wiring.
• MCP SDK + @supabase/mcp-server-supabase for tools.
• Sentry for prompt error tracking.

Entry points for modification

• Tweak a prompt → packages/ai-commands/src/.
• Add a new MCP tool → register a tool that calls existing data hooks; expose it through the Assistant.
• Tune evaluation → apps/studio/evals/assistant.eval.ts.
• Inline AI in a different surface → reuse the data/ai/ hooks; mount a streaming UI component.