AI Assistant
Built-in AI for writing, explaining, optimizing, and fixing SQL. One active provider drives every AI feature: chat, inline suggestions, editor actions, and fix-error.
The chat panel has three modes (Ask, Edit, Agent). It can call tools to inspect your schema, take context from @ mentions, and run templates from / slash commands. Provider keys stay in the macOS Keychain. Per-connection rules and tool whitelists sync via iCloud with the connection.
Open Settings (Cmd+,) > AI. The tab is modeled on Xcode’s Intelligence settings: a single Active Provider picker at the top, a list of configured providers below, and a detail sheet when you drill into one.
Add a Provider
- Click Add Provider… and pick a type: GitHub Copilot, ChatGPT, Cursor, Claude, OpenAI, OpenRouter, OpenCode Zen, Gemini, Ollama, or a custom OpenAI-compatible endpoint.
- Enter the API key, or sign in for Copilot and ChatGPT.
- Enter a model name, or pick one from the fetched list. Click Reload if needed.
- Click Test Connection.
API keys are stored in the macOS Keychain. Ollama is detected at launch.
Active Provider
The active provider handles every AI request. Override it per turn from the model picker in the chat composer. Change the default from the Active Provider picker at the top of the AI tab.
GitHub Copilot
Add Copilot like any other provider. The detail sheet runs GitHub’s device-flow sign-in: enter the displayed code on github.com to authorize. The Copilot language server starts when you add a Copilot provider and stops when you remove it.
Copilot supports tool calling through GitHub’s conversation/registerTools bridge. Tool calls go through the same approval flow as the other providers.
ChatGPT
Sign in with your ChatGPT account to use the Codex quota included with Plus, Pro, Business, and Enterprise plans, with no API key. The detail sheet opens your browser to sign in, then captures the redirect on a local callback. Tokens are stored in the macOS Keychain and refreshed automatically. If you already use the Codex CLI, click Import from Codex CLI to reuse that login.
The active provider then serves chat and inline suggestions from the subscription models (GPT-5.5, GPT-5.4, GPT-5.4 Mini). This uses an unofficial OpenAI interface that may change, and access follows OpenAI’s terms.
Cursor
Use your Cursor subscription for AI chat and inline suggestions, two ways:
- API key: paste a key from the Cursor dashboard. Stored in the macOS Keychain like other provider keys. Calls Cursor’s Cloud Agents API.
- Sign in with the Cursor CLI: leave the key blank and click Sign in with Cursor. This needs the Cursor CLI installed (
curl https://cursor.com/install -fsS | bash); the button runs agent login in your browser, and chat runs through the local agent command. No key to paste.
Pick a model such as Composer 2, or one of the frontier models Cursor proxies.
Cursor runs as an agent rather than a chat-completions endpoint. Two limits follow: the AI cannot call TablePro’s database tools (chat answers from the schema context TablePro sends, but Edit and Agent modes do not run queries through Cursor), and responses can be slower than the other providers. Using the API or CLI in your own app is permitted by Cursor’s terms.
Chat
Press Cmd+Shift+L, click the inspector toggle and pick AI Chat, or use View > Toggle AI Chat. The right inspector has a Details / AI Chat segmented picker at the top.
Type a question and press Return. Code blocks have Copy and Insert buttons. Insert opens the SQL in a new query tab, or fills the current editor if it’s empty. Token counts show below each response.
Conversations auto-save and auto-title from your first message. The composer footer’s clock icon opens recent conversations and the pencil-and-square icon starts a new one.
Failed responses show Retry. Successful ones show Regenerate. Click Stop to cancel a streaming response.
Chat Modes (Ask / Edit / Agent)
The mode picker in the composer footer controls which tools the AI can call. The mode sticks across turns until you change it. New chats start in Ask.
| Mode | Tools available | When to use |
|---|
| Ask | Read-only schema lookups: list_connections, get_connection_status, list_databases, list_schemas, list_tables, describe_table, get_table_ddl. | Asking questions, exploring a database, drafting queries you’ll run yourself. |
| Edit | All Ask tools, plus execute_query for SELECT, INSERT, UPDATE, DELETE. Destructive DDL (DROP, TRUNCATE, ALTER...DROP) stays blocked. | Letting the AI run queries it proposes, e.g. “fetch the 5 most recent orders”. |
| Agent | All tools, plus confirm_destructive_operation for destructive DDL. Runs tools in a loop, up to 10 round trips per turn. | Multi-step migrations, schema changes, larger refactors. |
- Run: approve this single call.
- Always for this connection: approve and add the tool to this connection’s whitelist (
Connection.aiAlwaysAllowedTools). Future calls of the same tool on this connection skip the prompt.
- Cancel: reject the call. The model gets an error result and continues.
Whitelists are per-tool and per-connection. They stick with the connection and sync via iCloud.
If the connection’s safe-mode level is Silent, write tools auto-approve without prompting. If safe mode is Read Only, write tools auto-deny and the message tells you to raise the level.
The tool call card expands to show the arguments and the response. Once the model is done calling tools, it streams the answer.
The cap is 10 tool round trips per turn. If you hit it, send a follow-up to continue.
| Tool | Purpose | Available in |
|---|
list_connections | List configured connections | Ask, Edit, Agent |
get_connection_status | Connection health and metadata | Ask, Edit, Agent |
list_databases | Databases on the active connection | Ask, Edit, Agent |
list_schemas | Schemas in the active database | Ask, Edit, Agent |
list_tables | Tables in a schema | Ask, Edit, Agent |
describe_table | Columns, types, constraints | Ask, Edit, Agent |
get_table_ddl | CREATE TABLE statement | Ask, Edit, Agent |
execute_query | Run SELECT / INSERT / UPDATE / DELETE. Multi-statement input rejected. Destructive DDL blocked. | Edit, Agent |
confirm_destructive_operation | Run destructive DDL after the model passes the verbatim phrase I understand this is irreversible | Agent |
Attach Context with @
Type @ in the composer to anchor a picker at the caret. Pick from:
- Schema: every table’s columns and foreign keys.
- A specific Table: that table’s columns and foreign keys only.
- Current Query: whatever’s in the active editor tab.
- Query Results: the most recent query result snapshot.
- Saved Query: pick one of your starred queries to send its name and SQL alongside the message.
Up/Down navigates, Return or Tab inserts, Escape dismisses. The @ button next to the composer opens the same picker as a menu.
Attached items show as chips in the composer. Saved query chips resolve when you send, so the AI sees whatever’s in the saved query at send time, not when you attached it.
Slash Commands
Type / (or click the ⌘ button next to the composer) to run a command. Built-ins:
/explain: explain the active query./optimize: suggest optimizations for the active query./fix: fix the last error against the active query./help: list the commands inline in the chat.
Custom Slash Commands
Add your own under Settings > AI > Custom Slash Commands. A command needs a name, an optional description, and a prompt template. Templates take these placeholders, substituted at send time:
{{query}}: the current editor query.{{schema}}: the formatted schema for the active connection (capped by Max schema tables in AI settings).{{database}}: the active database name.{{body}}: text typed after the command. For example, /review WHERE clauses passes WHERE clauses as {{body}}.
Save needs a name and a non-empty template. Saved commands show up in the slash menu next to the built-ins.
Inline Model Picker
Next to the mode picker is a model picker (cpu icon) listing every configured provider and its models. Pick one to override the active provider for the current turn. The label shows your pick; clear it to fall back to the active provider.
The override is per turn, not per chat. The next turn starts from the active provider unless you pick again.
Per-Connection AI Rules
Pin context to a connection so the AI sees it on every chat turn against that database. Open the connection’s edit form and pick AI Rules in the sidebar (sparkles icon).
Rules are plain text. Use them for facts the schema does not show:
- Table conventions:
Tables prefixed with tmp_ are scratch and safe to ignore.
- Join keys:
users.email_hash is the join key, not users.email.
- Soft deletes:
Always filter orders by deleted_at IS NULL.
- PII to avoid:
Never select users.ssn or users.dob.
- Business rules:
Active accounts have status = 'active' AND verified_at IS NOT NULL.
The text is appended to the system prompt under a ## Connection-Specific Rules heading, after the schema and any attached context. Rules sync via iCloud, so they apply on every device tied to the connection.
Inline Suggestions
Toggle Enable inline suggestions while typing in the AI tab. The active provider drives suggestions. Press Tab to accept, Esc to dismiss.
With Copilot active, suggestions come from Copilot’s inline-completion model. With any other active provider, suggestions come from chat completions.
Editor Actions
Right-click SQL in the editor for Explain with AI and Optimize with AI.
Default shortcuts:
| Action | Shortcut |
|---|
| Toggle AI Chat | Cmd+Shift+L |
| Explain with AI | Cmd+L |
| Optimize with AI | Cmd+Option+L |
Ask AI to Fix
Query error dialogs include an Ask AI to Fix button. It opens chat with the failed query and error pre-filled.
Context
Under Settings > AI > Context:
- Include database schema (default off)
- Include current query (default off)
- Include query results (default off)
- Max schema tables (default 20)
New installs default these toggles to off. The AI then only sees what you attach with @ mentions or what tool calls pull in. Turn the toggles on to auto-include context on every turn. Existing installs keep their previous values.
Privacy
Set a per-connection AI policy in the connection form: Use Default, Always Allow, Ask Each Time, or Never. New connections default to Ask Each Time.
External AI Clients
External clients (Raycast, Cursor, Claude Desktop, and other MCP clients) call the same AI tools through the External API. Two per-connection settings gate them:
- AI policy decides whether the connection is reachable by AI clients at all.
Never blocks every external AI tool call against this connection.
- External Access caps the level:
blocked, readOnly (default), or readWrite. A token’s effective permission is MIN(token.scope, connection.externalAccess). Set this in the connection form’s Advanced tab.
See Tokens for the scope model.
