Everything you need to start, configure, and manage chats with CellCog agents: the editor chips, modes and tiers, the header bar, per-message actions, chat search, and the Agent Computer.
Cowork, Browse, and Tools each carry an approval level for the chat, so you decide how much the agent may do on its own before asking you.
The Tools chip has three positions:
You can connect a new tool from inside the picker, and the gear icon takes you to your connections. AI employees keep their own Tools selection, separate from your chats. See the Connectors & Tools Guide.
Browse works through your own Chrome and the CellCog extension. The chip lists your Chrome profiles: a green mark means the extension is installed and connected in that profile, amber means installed but not connected right now, and gray means the extension is not installed there. Click a profile to enable Browse for this chat with that profile; commands route to the profile the chat is bound to, so you can keep several profiles open.
AI employees are different: each one has its own cloud browser with its own logins, managed from Cloud Browsers (in the sidebar's More menu). Cloud browsers belong to employees, not to chats. See the AI Employees Guide.
You can attach files directly to any chat message. Each file can be up to 1GB; large uploads use resumable transfer, so a flaky connection picks up where it left off instead of starting over. Images are limited to 25MB per file. Uploaded files land in the agent's workspace, where it can read, analyze, and transform them.
Every chat runs on two settings you pick together in the selector: a mode (which agent does the work) and a tier (how much depth it runs with: Flash · Core · Max). Any combination is one click, and every mode row carries an info icon that opens a learn card with Speed and Depth meters, so the tradeoff is visible before you pick.
Best for:
Think of it as: one agent that executes your instructions, iterates on feedback, and produces videos, images, reports, code, and more, across every modality rather than just code.
When to avoid: when you need deep, multi-angled research out of the box, like academic work. Agent Team does that by default.
Best for:
Think of it as: our most imaginative agent, the same single agent as Agent Mode, tuned for taste. Give it a design brief and it delivers polished, opinionated results.
When to avoid: programmatic work, coding-heavy tasks, or deep research. Use Agent Mode for fast iterative and programmatic work and Agent Team for deep research.
Best for:
Think of it as: a team of specialized agents that research deeply, challenge each other's findings, and produce well-rounded outputs in any format you need.
When to avoid: when you expect many iterations to reach the final output. Each run costs more, so for back-and-forth refinement, Agent Mode gives more control at lower cost per iteration.
| Tier | What it means |
|---|---|
| Flash | Our quickest, most economical configuration of that agent. Answers begin almost immediately. Good for everyday questions and edits. |
| Core | The balanced everyday point, with real reasoning depth. |
| Max | The deepest configuration we can build for that agent. |
The selector stays live inside a conversation: switch between Agent Mode and Agent Creative, or change any mode's tier, at any time. Changes apply from your next message, with no lost context. The one exception is Agent Team: its multi-agent setup is created when the chat is born, so Team is chosen at creation and cannot be switched on or off mid-conversation.
It became a cell on the grid: Agent Team on the Max tier, every setting maximised, for high-stakes work like legal analysis and financial decisions where quality justifies the extra cost. Same premium orchestration as before, and existing chats keep their exact behavior.
The modes above are how you work with a CellCog agent directly: you send a message, it responds. If you instead want a worker that runs on its own, hire an AI Employee. An AI employee runs on the same engine, but it has its own email inbox, its own cloud browser, a place on your organization's task board, routines that wake it on a schedule you or it defines, and a memory that persists across sessions. You can also hand it work from any chat by asking the agent to put a task on that employee's board. See the AI Employees Guide.
In an active chat, the header shows the chat's mode and tier (hover the icon for a description), a meter of how much of the conversation's context is in use, and these controls:
| Control | Function |
|---|---|
| Agent Computer | Open the agent's filesystem and workspace (in an AI employee's chat this is the Employee Computer, which opens on its Home app) |
| Save Artifacts | Save generated files to My Drive, the Organization Drive, or a project |
| Update Agent Memory | Store this chat's insights in the agent role's memory (shown only when the chat has an agent role) |
| Share | Copy a link to the chat, make its apps public or private, or publish a community post |
| View Costs | See how many credits this chat has used |
Every message carries a small toolbar that appears when you hover over it: the timestamp and Copy message on every message, plus Save Artifacts and Share on the latest agent reply. On touch devices there is no hover, so the buttons stay visible. Links to apps the agent built appear as tiles inside the message itself.
Agents can write math and it renders in chat: $...$ for inline expressions and $$...$$ for display equations. Plain dollar amounts such as "$25 and $30" stay ordinary text; rendering only switches on when the span contains real math notation.
Click the magnifier at the top of the sidebar (or press ⌘K / Ctrl+K) and type at least two characters. Search covers chat names and message content across all your chats, including your AI employees' sessions, and also matches your employees, channels, and DMs by name. Name matches come first, then content matches with highlighted snippets showing why a chat matched. Use the type filter to narrow the results (for example to Employees), then click a result to open it.
The sidebar shows a shortcut next to New chat when you hover it, in your platform's keys. On a Mac the modifier is ⌘; on Windows and Linux it is Ctrl.
| Action | Mac | Windows / Linux |
|---|---|---|
| New chat | ⇧⌘O | Ctrl+Shift+O |
| Search chats | ⌘K | Ctrl+K |
| Show or hide the sidebar | ⌘\ | Ctrl+\ |
| Start or stop voice input | ⇧⌘U | Ctrl+Shift+U |
| Send message | Enter | Enter |
| New line in the editor | Shift+Enter | Shift+Enter |
The sidebar groups your work into AI Employees and Chats. Each chat row shows:
Hover a chat and click its three-dot menu (⋮) to Rename it, Add to project (or Move to project and Remove from project once it is in one), or Delete it. Chats can be filtered by project using the project selector in the sidebar navigation. The sidebar's More menu holds Drives, Avatars, and Cloud Browsers, plus the Learn links (Podcast, Blogs, Tutorials) and Pricing.
⚠️ Warning: deleting a chat permanently removes all messages and history, generated files, code, and apps, associated media, and uploaded files. This action cannot be undone.
The Agent Computer lets you see the agent's workspace:
The Agent Computer is a file browser. It does not include a web browser tab; only AI employees have a browser of their own (the Browser app inside the Employee Computer). Files stay put between sessions: workspaces of chats you have not opened in a while are archived and restored automatically the next time you return, and software the agent installs in its workspace persists with it.
Save generated content to persistent storage:
Apps the agent built can be saved as .ccapp shortcuts that open the app straight from Drive. See the Drive Guide and the Apps Guide.
For project chats with an agent role, you can manually update the agent's memory:
Requirements:
The same memory update also runs automatically after the chat has been idle for a while. See the Memory System & Context Trees Guide.
When an agent builds an HTML app, dashboard, or report, it lives in the chat's workspace and is served at its own URL. Apps are private by default; the Share control makes them public. Paste an app's URL into a new chat to keep editing it. See the Apps Guide.
CellCog agents excel at complex, multi-step work described in natural language. You do not need special syntax or a builder: just describe what you need.
Write your prompt as a natural sequence of steps. The agent executes them in order, using earlier results to inform later steps. For example:
"First, research the top 10 competitors in the AI coding assistant space. Then analyze their pricing models and create a comparison spreadsheet. Finally, based on that analysis, write a strategic recommendation for our pricing."
The agent will:
CellCog's average first prompt is long. Our users routinely submit multi-step jobs with multiple deliverables in a single message, and executing a complex, multi-deliverable brief end-to-end (reports, spreadsheets, presentations, images) in one go is one of CellCog's core strengths.
For work that should happen on a schedule without you starting it, use an AI employee: put standing work on its task board and give it a routine. Chats are for interactive work you start yourself.
Chats are named automatically from their content. To choose your own name, hover the chat in the sidebar, open its three-dot menu, and select Rename.
Two ways, same result: open the chat's three-dot menu (⋮) in the sidebar and pick Add to project, or open the chat and choose the project in the Project chip of the message editor. From then on the chat can read the project's documents, and it appears on the project's Chats tab. Pick Remove from project on the same menu to unlink it.
Three ways. A sound plays in the tab when a chat finishes, needs you, or stops with an error (mute it under Settings, Notifications). For a system notification on your devices even when the tab is closed, turn on Browser notifications in Settings, or accept the offer that appears once a chat has been working for a moment: the switch is for your account, and each device asks once for your browser's permission (the same Settings row shows whether the device you are on has allowed them). Sound and browser notifications follow one rule: they fire when a chat finishes, needs you, or stops with an error, and only for the person who sent the message that started the work. An AI employee working on its own schedule does not notify anyone; what it needs from you appears on your task board and in its report. And if a completed chat stays unopened for 30 minutes you get a reminder email, at most a few per day. The email says whether the chat finished or is waiting on your input, and matches how the chat was started: opened in the browser, started by your agent through an API key (with the retrieval snippet), or started by a connected app such as Claude (with what to tell it). A result your agent or connected app already fetched counts as seen, so those chats never trigger the email.
No. Chat deletion is permanent and cannot be undone. Save important artifacts before deleting.
Click any chat in the sidebar, or find it with Search chats. Your full history is preserved.
Open the header menu and select View Costs.
No. Scheduling belongs to AI employees: routines wake an employee at the times you or it define. See the AI Employees Guide.