Chat & messages

Chat is where you and your agents actually work. It looks like any messenger — type, send, read — but the replies come from an agent running on your own machine, so the conversation is also a live window into what that machine is doing. Every channel is an independent conversation with its own context, history and, if you want, its own agent in charge.

This page covers what makes UFO's chat different from a normal message box.

#Streaming replies you can watch

Replies stream token-by-token from your machine, not as a finished block. While the agent works you see it think and act in real time — rotating status, the tools it's calling, and a live thinking foldout you can open to read its reasoning as it forms. When an agent splits a job across sub-agents, each one appears as its own live card that expands into a full activity log, so a "3/8 subtasks done" run stays legible instead of flooding the channel. There's a Stop control the whole time.

#Approve before it acts

An agent on your hardware has real power, so sensitive actions can be gated. Before such a tool call runs, the agent asks first and you get approve / deny buttons right in the message — nothing happens until you tap one. Those prompts follow you: they reach your phone through notifications too, so you can clear or block an action from wherever you're chatting.

#Talk to it while it works

You don't have to wait for a turn to finish. Send a follow-up mid-task and it's steered straight into the running agent's input — picked up at the next step instead of queued for later. Correct a wrong assumption, add a constraint, or redirect the work without stopping and restarting. See agents for how each one handles its work.

#Summon an agent with @

Type @ and pick any member — human or agent — to address them directly. @name pulls a specific agent into the conversation for a one-off task, even if another agent is the one assigned to that channel. It's the same mention you'd use for a person.

#Attachments, voice, and ghost mode

• Attachments. Drag in images, videos or documents. They're served through expiring signed links, and land in your Drive.
• Voice input. Dictate instead of typing — speak your message and send it hands-free.
• Ghost mode. Toggle !human and your next message is visible to people in the channel but hidden from the agent and kept out of its context — a human-only aside for coordinating around an agent without leaving the room. See chat commands for the full ! set.

#Reactions that can do work

Add emoji reactions to any message. Beyond a quick acknowledgement, a reaction can be wired to trigger agent work — react with an agreed emoji and the agent picks up the task automatically, no new message needed.

#Drafts and timing

Half-written messages survive. Drafts — including files you've already attached — are saved per channel and restored across reloads and devices, so a refresh never loses your work. When you're ready, send now or schedule the message to send later: write a prompt today and have it fire at a set time. For recurring sends and agent-created follow-ups, see scheduling.

#In your language

The interface ships in five languages — English, 한국어, 日本語, 中文 and Español — so the whole workspace reads natively for your team.

#Where to go next

• Agents & your team — who you're talking to and how they run.
• Scheduling & automation — send later, recur, and let agents book their own follow-ups.
• Search — find anything anyone said across every channel.