Chat

The chat interface is where you collaborate with your AI agent. Everyone in your workspace can participate in any implementation's chat, making it easy for your team to review progress, ask questions, and provide feedback.

Sending messages

Type your message and click the send button to send it to the agent. You can also press Cmd/Ctrl+Enter to submit from the keyboard, along with the other keyboard shortcuts. The agent will respond and continue working based on your input.

You can send messages even while the agent is still working. When you're running Claude Code, your message reaches the agent mid-turn so it can change course right away — you'll know steering is live because the placeholder changes to Message the running agent.... See Steering a running agent for more details. For all other agents, your message is queued and picked up automatically when the agent finishes its current task.

If chat shows a warning that the agent cannot reach a domain, see how to choose what to allow and its scope. After you change network access, send a message to start the agent's next run with the new rules.

Quoting text from the chat

Select any text in the chat and an Add to Message button appears next to your selection. Click it to drop that text into the message box as a quote. This is the fastest way to point at one specific thing — a claim the agent made, a file path it mentioned, a line of its work log — instead of describing it.

You can quote from agent messages, your own messages, and the agent's work log. Quote more than once before sending and each block is appended, separated by a divider.

To point at code rather than chat text, comment on the diff instead.

Stopping the agent

If the agent is heading in the wrong direction or doing something you don't want, click the Stop button in the chat input area to halt its current work. You can then send a new message to redirect the agent.

With Claude Code, you often don't need to stop first — send your correction while it's still running and it will adjust without losing the work it has already done. Stop when you want to abandon the current attempt outright.

Viewing the agent log

To see exactly what the agent did, expand the Show Work dropdown next to the agent's name in the chat. This shows a detailed log of the agent's work, including tool calls and intermediate steps.

Running shell commands

Type ! in the chat input to toggle into shell mode. The input field switches to a dark theme with a "Shell command..." placeholder, indicating that your next message will be executed as a shell command on the cloud instance rather than sent to the agent.

This is a quick way to run commands—like checking logs, running tests, or inspecting files—without switching to the Terminal. Press Escape to exit shell mode and return to normal chat.

Uploading files

Chat supports several ways to share files with your agent:

• Drag and drop files directly onto the chat area
• Paste images from your clipboard (Cmd/Ctrl+V)
• Click the attachment button to browse and select files

You can upload any file type — all uploads are shared with your agent. Images and videos preview inline in the conversation; everything else attaches as a file you can download:

• Images (up to 10 MB): JPEG, PNG, GIF, WebP, BMP, HEIC, SVG — displayed inline
• Videos (up to 100 MB): MP4, MOV, WebM, AVI, MPEG, WMV, FLV, 3GP — playable inline
• Documents (up to 25 MB): PDF, Word (DOCX), Markdown, plain text
• Code files (up to 25 MB): JavaScript, TypeScript, Python, Ruby, Go, Rust, and many more
• Data files (up to 25 MB): JSON, YAML, CSV, XML

All other file types (up to 25 MB) are also accepted and shared with the agent.

PDFs and Word documents that agents save as artifacts do get a full in-app viewer, with page navigation and zoom.

Collaborating with teammates

The avatars at the top of an implementation show teammates who follow the implementation, have participated in its chat, or are currently viewing it. Up to three avatars appear in the header; open the ⋯ menu to see additional teammates.

Presence indicators show who is currently in the implementation:

• A green dot means the teammate is actively viewing it.
• A gray dot means the teammate still has it open, but their tab is in the background.
• No dot means the teammate is not currently viewing it.

When a teammate starts writing a message, you'll also see their name and is typing… above the chat input.

Teammates who join the chat do not change which agent credentials the implementation uses.

To keep up with a shared implementation, follow it. You can change the image teammates see for you under Account Settings → Profile.

Tips for getting the most from your agent

Ask for artifacts

Request screenshots, diagrams, or other visual outputs to document progress. Agents can share these directly in chat without committing them to the repo. See Artifacts for more details.

Iterate from Slack and GitHub

Continue conversations from Slack by replying in the same thread, or from GitHub by commenting on PRs. Your feedback flows back to the agent automatically, and the agent will notify you once it's done.

Use MCP Playwright for in-app testing

If Playwright MCP is enabled for your project, ask the agent to test code changes directly in the browser and take screenshots. This lets the agent verify UI changes work correctly without you having to run the app yourself.

Take advantage of Sentry integration

With Sentry MCP enabled, agents can directly query your Sentry account to understand what errors are happening in production. Agents can view actual production errors and their surrounding context, helping them write better fixes.

Comment on specific lines of code

When reviewing the diff, click the + on any line to add a comment. Your comment is pasted in the agent chat with full context about the code you're referencing, making it easy to request precise changes.