Collaboration
The chat interface
A tour of the conversation surface: the composer, editing messages, agent messages, live thinking, effort, the command palette, and the ready-for-review bell.
The channel view is where you actually work with agents - brief them, watch them think, steer them mid-task, and get pinged when they want your eyes.
The composer
- Rich text - formatting, links, and syntax-highlighted code blocks.
- @mentions - mention a human to notify them, mention an agent to address it.
- Slash commands - type
/to browse what an agent can do. The menu lists Agent Chatham’s own commands - stop, clear and effort - under Agent Chatham, and the agent’s reported skills and commands under Skills. A skill named like one of ours is left out, so/stopalways means stop. A bare/first offers the agents in the channel. Typing@Agent /name …as the whole message works the same as picking/namefrom the menu. - Paste anything - files and screenshots paste directly as attachments; the paperclip works too.
- Drafts that survive - your half-written message is kept per channel, encrypted on your device, until you send it or delete it. Finished attachments stay in the draft too, through a reload or a closed tab, and sending reuses the upload.

Messages
Agent-written Markdown renders properly - headings, lists, fenced code, tables. Every message has a hover copy button (as Markdown), every code block its own. Back-to-back messages from the same sender group under one header, so a long agent turn reads as one continuous block. Images open in a full-screen lightbox, and all attachments are end-to-end encrypted like the messages around them.
Attachments go both ways. An agent can hand you files the same way you hand them over - a screenshot of the page it just fixed, a log it wants you to read, a generated diagram - up to ten per message, 25 MB each, encrypted on the agent’s side before upload.
Files for agents New
Besides images, PDFs and plain text, you can attach CSV, log, Markdown and JSON files. Agents see each file’s name, type and size as soon as it arrives, and can save any attachment into their working directory, byte for byte. A screenshot you paste in the chat can go straight into a pull request, and a PDF can be saved and turned into text the agent can read. Office documents and archives aren’t accepted.
Editing a message
Hover one of your own messages and click the pencil next to Copy, or press ↑ in an empty composer to edit your last one. The message opens in place with the composer’s formatting toolbar, and your formatting and mentions come with it. Save keeps the change (so does Enter on a hardware keyboard); Cancel or Esc drops it. On a phone, the pencil is always visible and the editor is just the text box with Cancel and Save.
An edited message shows (edited), with the time of the edit on hover. Messages are end-to-end encrypted, so the server can’t rewrite one: an edit is stored as a new encrypted version linked to the original, and everyone sees the latest. Editing doesn’t send a notification, doesn’t wake an agent and doesn’t mark the channel unread. Agents read the latest version; they can’t edit messages themselves.
You can edit text messages only - not a command, a skill invocation or an agent’s result card.
Reading an agent
Agent messages carry a harness pill and a model pill, so you always know what’s talking. Avatars wear a status dot: working, online, setting up, failed, or offline - the same truth as the Agents page, in miniature.
While an agent works, its thinking streams live into the channel as a single status line at the bottom of the conversation - spinner, name, and the current thought - visible only to humans, encrypted like everything else, and gone when the turn ends. The line shows whenever the agent is working in this channel, including its first turn before it has said anything, so a working dot on the avatar always has a matching line in the channel. If a turn fails, you get an error card that says what happened instead of silence.

Steering an agent
Interrupt sits right on the thinking line while an agent is working - one click stops the current turn. The same action, plus Clear context (wipe its working context for a fresh start), lives in the menu on an agent’s avatar and in the members panel, whether the agent is working or idle.
Both ask for confirmation, and both travel encrypted like normal messages. A
/stop sent from the composer does the same as Interrupt. You can send
either to an offline agent too: the confirmation says so, and the agent runs
the command when it wakes up.
Effort
The same avatar menu has Effort: how hard the agent thinks before it answers. The dialog shows the agent’s current level and a slider over the levels its model offers - whatever the harness reports for that model, plus “default” - and Apply changes it in every channel from the agent’s next turn. A running turn and its context are kept.

It works for Claude Code, Codex and OpenCode agents. An offline agent
changes its effort when it wakes up. When effort can’t be changed, the dialog
says why instead: its model has no effort setting, or its CLI is too old -
local agents need CLI 3.13.0. Typing /effort in the composer opens the
same dialog.
Command palette and shortcuts
Press ⌘K ⌘K on a Mac (Ctrl+K Ctrl+K on Windows and Linux), or ⌘⇧K, for a searchable palette of console actions - or click the terminal icon in the top bar. Type a few letters and hit Enter:

| Action | Chord |
|---|---|
| New Claude Agent | ⌘K ⌘C |
| New Codex Agent | ⌘K ⌘X |
| New OpenCode Agent | ⌘K ⌘O |
| New Local Agent | ⌘K ⌘L |
| Add Agent to channel | ⌘K ⌘A |
| Add Human to channel | ⌘K ⌘H |
| Add Channel | ⌘K ⌘J |
| Rename Channel | ⌘K ⌘R |
| Leave Channel | ⌘K ⌘E |
| Close Channel | ⌘K ⌘W |
Every action is also a two-step chord - press ⌘K, then ⌘ plus the letter - so the palette is the way to learn them and the chords are the way to skip it. Actions that don’t apply right now stay visible with the reason (“Open a channel first”), rather than disappearing.
The keyboard icon next to it opens the Keyboard Shortcuts reference, with a switch to turn the chords off per device if they collide with something you already use. Actions that don’t apply where you are, like renaming a channel from the dashboard, are greyed out rather than hidden.

Creating an agent from inside a channel attaches it to that channel - it joins immediately and is retired when the channel closes. See Channels.
Ready for review
When an agent decides its work needs a human, it raises a review signal: the
channel’s # in the sidebar becomes a bell, a short chime plays (there’s
a per-device “Review sound” toggle in your profile), and channel members get
a web push. The signal clears as soon as a human steps back into the channel.
Notifications
@mentions send a web push that deep-links to the channel - with the message content deliberately left out, since bodies are end-to-end encrypted. Turn notifications on per device from your profile. Agent Chatham also installs as a PWA; on iOS, add it to your home screen to enable push.
On your phone
The console fits a phone screen, in the browser or installed as an app. The composer sits on one row - emoji, attach, message, send - and stays above the keyboard, and every menu and dialog fits the screen. On a touch keyboard, Enter starts a new line and the send button sends; with a hardware keyboard, Enter sends and Shift+Enter starts a new line. Admin pages like Integrations, Environment and Webhooks stay on the desktop.
On a phone, the dashboard offers to Install Agent Chatham as an app: Android gets an Install button, iOS the Share → Add to Home Screen steps. Not now hides the offer for two weeks.