Chats overview

Multi-turn conversations with your data and your codebase, powered by frontier models.

Last updated

Chats are where you talk to a model with rich context: a connection’s schema, a result you’re staring at, a snippet you’re trying to fix, or a thread you started yesterday.

Start a chat

  • From the sidebar — New chat → pick a model (or stick with the default prisma).
  • From the editor — Right-click a query → “Discuss in chat”. The query and any visible result are attached.
  • From a result — Right-click a row → “Ask about this row”. The row JSON becomes the chat’s first attachment.

What gets attached automatically

SourceAttached
New chatActive connection’s schema digest
From editorSchema + current SQL + last result rows (capped)
From a rowSchema + the row JSON
From an errorSchema + statement + error message

You can detach anything from the Context panel on the right. If a teammate later opens the chat, they see the same attachments — but only if they have access to those connections / snippets.

Model picker

The dropdown at the top of the chat lets you change models mid-thread. The new model gets the full conversation history.

Switch reasons:

  • Stuck with prisma? → switch to tambor, ask the same question.
  • Long thread getting expensive? → switch to haiku for follow-ups that don’t need depth.

Tool calls

Models in chats can call tools — primarily the MCP SQL tools. When a model proposes a tool call, Cocobox shows it inline:

🔧 sql_run on prod-readonly — SELECT count(*) FROM users WHERE … [Approve] [Edit] [Reject]

You stay in the loop. Approval is required by default for any write statement; auto-approve for read statements is configurable per workspace.

What’s next