Chat

Chat is the primary interface for interacting with AI in Cosmo. Send messages, ask questions, generate artifacts, and use your connected tools — all through natural language.

How Chat Works

When you send a message, Cosmo processes your request through its AI engine. Depending on the request, it may:

Respond with text directly
Call one or more tools (integrations, task management, artifact creation)
Create or update artifacts in the Hub
Manage your workspaces and tasks

Responses stream in real-time so you see results as they're generated, not after a long wait.

Live Activity Indicators

Cosmo shows you what's happening at every stage so the chat never feels stuck.

When Cosmo is processing your request — before the first response token arrives or between tool calls — you'll see a subtle "Thinking" indicator with animated dots. This tells you the AI is working, not frozen.

Tool Calls

When Cosmo uses a tool (like creating a task, searching Notion, or updating an artifact), you'll see an activity badge showing the tool name and a spinner. For example:

Tasks: Create — creating a new task
Workspaces: List — checking existing workspaces
Notion: Search — searching your connected Notion workspace

Once a tool finishes, the spinner changes to a checkmark. You can click the badge to expand and see a preview of the tool's result.

Streaming Text

As Cosmo generates its response, text streams in word-by-word with a blinking cursor. You can read the response as it forms rather than waiting for it to finish.

Code Blocks

When a response includes a fenced code snippet, it renders in its own panel with the language shown at the top. Click Copy in the panel's header to copy the exact code to your clipboard — the button briefly confirms with a checkmark once it's copied.

Long Prompts

When a message you send runs longer than about five lines (or roughly 350 characters), the chat shows only its opening, ending in "…", with a Show more link underneath. Click it to see the full prompt; click Show less to fold it back. Short messages are always shown in full.

This only changes what is displayed — Cosmo always receives the complete prompt, and expanding or collapsing it never affects the response. Long prompts start collapsed again whenever you reopen or reload a chat. Line breaks, lists and code are kept as you typed them in both views.

Multi-Step Responses

Complex requests often involve multiple steps. Instead of showing one long message at the end, Cosmo breaks its response into visual segments:

First text block — Cosmo explains what it's about to do
Tool activity — badges show tools being called (e.g., listing projects, creating an artifact)
More text — Cosmo summarizes what happened or asks follow-up questions

This makes it easy to follow along with multi-step operations like "create a workspace, add some tasks, and generate a proposal."

Answering Multiple Questions at Once

When Cosmo needs more than one piece of information before it can continue, it asks for all of it in a single card instead of one question per message.

The card shows one question at a time, with a counter (e.g. "2 of 4") so you know how much is left. For a question where more than one answer applies, you can select multiple options — the same one-click toggle pattern used for suggested next steps. If none of the offered options fit, use Other… to type your own answer — it can run over several lines, and the line breaks are kept in what Cosmo receives. Cosmo leaves catch-all choices such as "Other" or "None of these" off its own list of options, since Other… is always there.

You can skip a question you don't want to answer; Cosmo is told it was skipped rather than left to guess whether the answer was lost. Once every question is answered or skipped, submitting sends all of your answers back in a single message — you don't wait through a separate round trip for each one.

@-Mentions

Type @ in the message box to open a searchable list of files, workspaces, sub workspaces (folders), and people in your organization — similar to mentioning something in Slack or Notion.

Files — start typing a file's name to find it. Selecting it makes that file's content available to Cosmo for your current message, without attaching or re-uploading it.
Workspaces — mention a workspace to tell Cosmo which one your message is about for that turn.
Sub Workspaces — a folder inside a workspace. Mentioning one tells Cosmo the files filed under it are what your message is about for that turn.
People — mention a teammate to reference them in the message. Cosmo doesn't yet send them a notification when you do this; the mention is recorded in the message as a reference, visible to anyone who reads the conversation.

Keep typing after @ to narrow the list, use the arrow keys and Enter to pick a result with the keyboard, or click a result with the mouse. Press Escape to close the list without picking anything.

Web Search

Cosmo can search the web to answer questions about current events, look up facts, or find information that isn't in your connected tools.

Web search is built-in and always available — you don't need to configure anything. When Cosmo determines that a web search would help answer your question, it searches automatically and incorporates the results into its response.

Example prompts:

"What are the latest trends in AI-powered workspace management?"
"Look up the current pricing for Notion's team plan"
"What did Anthropic announce this week?"
"Find the best practices for running remote standups"

When a web search happens, you'll see a "Web Search" activity badge in the chat with a spinner, just like any other tool call. Once the search completes, the badge shows a checkmark and Cosmo uses the results to write its response.

What web search is good for

Current events and recent news
Fact-checking and looking up specific data
Finding documentation, guides, or best practices
Researching competitors, pricing, or market trends

What it's not for

Searching your own connected tools (use integrations like Notion, Google Drive instead)
Private or internal company information (web search only finds public data)

Command Permissions

Sometimes Cosmo needs to run a command on your system — for example, checking git status, installing a package, or listing files. Every command goes through a permission system that keeps you in control.

How It Works

Before any command runs, Cosmo shows you:

A plain-English explanation — what the command actually does, in everyday language (e.g., "List files in the current folder")
A risk badge — color-coded so you can decide at a glance:
- Safe (green) — read-only commands that don't change anything
- Moderate (yellow) — commands that modify files within your workspace
- Dangerous (red) — commands that affect things outside your workspace or have system-wide impact
The raw command — the exact command that will run, for transparency

You can Allow or Deny each command. If you deny, Cosmo stops immediately — it won't try to work around your decision or keep going with other commands.

What Gets Blocked Automatically

Some commands are too risky to even ask about. Cosmo blocks these silently and tells you it can't proceed:

Deleting system directories (/usr, /etc, /System)
Touching sensitive files (SSH keys, AWS credentials, GPG keys)
System-level operations (formatting disks, shutting down, rebooting)

These commands never reach you as a prompt — they're stopped before that.

Directory Sandboxing

Commands that modify files (copy, move, delete, create) are checked against your workspace folder. If a command tries to write or delete files outside your workspace directory, it's flagged as dangerous so you can make an informed decision.

For more details, see Command Security.

Context Window

Every AI model has a finite context window — the maximum amount of text it can process at once (conversation history, system instructions, and response combined). Cosmo tracks this for you.

Context Meter

The chat header shows a context meter next to the model name. It displays:

Current usage (e.g., "32k / 200k")
A color-coded progress bar: green (under 60%), yellow (60-80%), red (over 80%)

Hover over the meter to see exact token counts.

Auto-Compaction

When your conversation nears the context limit (80%), Cosmo automatically compacts older messages by summarizing them. This keeps the conversation going without losing important context.

What happens during compaction:

Recent messages (last 3 exchanges) are kept as-is
Older messages are summarized into a compact context block
The summary preserves key facts, decisions, and context

Compaction happens transparently — you don't need to do anything. Your conversation continues without interruption.

Model Switching

You can switch AI models mid-conversation using the model selector in the chat header. Click the current model name to open a dropdown with available models grouped by tier (Fast, Balanced, Premium).

Different models have different strengths:

Fast models (Haiku, GPT-4o Mini) — quick responses, lower cost, great for simple tasks
Balanced models (Sonnet, GPT-4o) — best for most work
Premium models (Opus) — maximum reasoning capability for complex tasks

See AI Providers for the full model list with context windows and capabilities.

Stopping a Response

If Cosmo is generating a response you no longer need, click the Stop button in the chat header. Whatever had been written so far is kept, marked "You stopped this response" so you can tell it apart from one that finished on its own.

When a Response Doesn't Finish

A response can end early for reasons that have nothing to do with you — the connection drops, a connector fails, your session expires — or because you stopped it, or because an action needed permission that wasn't granted.

Whatever the reason, Cosmo keeps the turn instead of discarding it:

Anything already written stays on screen, marked as incomplete rather than presented as a finished answer
The reason is shown underneath, so you can tell a network blip from a permission prompt you missed
A Retry button usually re-runs your original message — you don't retype it, re-attach files, or re-pick a model (see "What Retry does" below for the few failures where Retry isn't offered)

This all survives switching chats, restarting the app, and installing an app update — the conversation is saved on your machine, so it is there when the app comes back.

It also survives signing out and back in, and reinstalling on the same or a different machine. Those cases clear the local copy, so the conversation is restored from the cloud against your account instead. Either way an interrupted response is still there, still marked, still retryable when you come back to it.

One limit worth knowing: what is recovered is the conversation as last saved, not the half-written reply as it looked mid-stream. Cosmo saves a response when it finishes, so a reply cut off part-way comes back marked as interrupted with a Retry button rather than as partial text. The prompt is always kept, so nothing has to be retyped.

What Retry does

Retry re-sends the original message exactly as you wrote it, with the same attachments and the same model the first attempt used — not whatever model happens to be selected now. The failed attempt is replaced rather than stacked underneath, so retrying repeatedly doesn't fill the chat with dead ends.

Three failures are the exception, and Retry isn't offered for them: the conversation grew too large for the model's context window, the request itself was too large to send, or an attached image was too large to process. Resending the identical request would just reproduce the same failure, so you'll see a New Chat button instead, and the composer is locked for the rest of that conversation — starting a new chat is the only way to keep going.

If that attempt had already started writing an artifact, the unfinished draft is moved to Trash before the retry runs, so you get one finished document instead of two half-written ones. You can restore it from Trash during the grace period if you want it back.

Artifacts are the only thing Cosmo can take back for you. If the interrupted attempt had already created a task, or done something through a connected tool — sent a message, drafted an email — that has already happened, and retrying may do it again. Check the interrupted response for what it managed to complete before retrying it.

Retry is offered on the most recent interrupted response. Once you send a new message, the conversation has moved on and re-running an older turn would answer a question that is no longer the last one asked.

Sessions

Each conversation is a session. Sessions persist across app restarts and maintain full conversation history.

New session — click the new chat button in the sidebar to start fresh
Switch sessions — click any session in the sidebar to resume a previous conversation
Delete sessions — right-click a session to delete it

Automatic Titles

After your first exchange, Cosmo generates a short title for the session (e.g., "ACME Proposal Review") so you can find it later.

Workspace Context

Sessions can be associated with a workspace. When Cosmo sets a workspace context, all tasks and artifacts created in that session are scoped to the workspace. You can see and change the workspace context through chat.

When a session is bound to a workspace — for example, because you have that workspace open in the Hub — its name appears as a small context chip just above the message box, so you always know which workspace you're currently working in. Cosmo is also told the workspace's name and ID as part of its context for that turn, so it can act on the right workspace when you refer to "this workspace" without naming it explicitly. Click the × on the chip to remove the workspace from the conversation's context; opening it again in the Hub restores it.

Artifact Context

Opening an artifact (a document, report, or other file) in the Hub shows its name as a second context chip next to the workspace one, above the message box. Unlike a plain visual label, this one is also part of what Cosmo knows: whichever artifact is open is sent along with your next message, so you can say "explain this" or "summarize this" without naming the file — Cosmo reads that exact document rather than asking which one you mean. Click the × on the chip to stop sending that artifact as context; it reappears if you reopen the artifact.

Tips for Effective Chat

Be specific — "Create a bar chart of Q1-Q4 revenue" works better than "Make a chart"
Reference workspaces by name — "Add this to the ACME workspace" helps Cosmo find the right place
Ask for modifications — Don't recreate artifacts from scratch. Say "Update the proposal to include a timeline section"
Use natural language — You don't need special syntax. Just describe what you need as you would to a colleague
Watch the activity indicators — They tell you exactly what Cosmo is doing, so you know if it's still working on your request

On this page