Chat

Chat is where you talk to your workers and watch them work. You type a request, the worker thinks, calls tools, and streams its reply; the right-hand pane shows exactly what it’s doing — searches, file edits, browser actions — as it happens.

How to get here

• Click + New chat (always starts a chat with 🎩 Theodor), or
• Click any conversation in the sidebar, or
• Open a Worker and press Start chat.

The composer (where you type)

• Message box — type your request and press Enter to send. You can keep typing while a worker is still replying; queued messages are auto-sent when it finishes.
• 🎤 Dictate — speak instead of type (voice-to-text).
• 🎧 Talk — a live, hands-free voice conversation (full-duplex), when the worker’s provider supports it. There’s a per-minute cost; see Voice settings.

Starter prompts & task templates

An empty chat offers two collapsible strips above the composer:

• Question examples — plain conversational openers (“What could I even ask?”). Click one to drop it into the composer.
• Task templates — a library of 116 ready-to-fire prompts across 19 categories: organizing files, analyzing data, building a knowledge base, checking system health, touring a codebase, business research, content production, automation routines, and more. Each template is filtered to the worker’s actual tools, so you only ever see what this worker can really do — plus a search box to find the right one fast. Clicking a template sends it immediately.

The conversation

Each turn shows the worker’s reply as it streams. Inline you may see:

• Tool-run cards — each web search, file read/write, or browser action the worker took, with its result. A 🔁 chip marks a cached tool result (reused to save tokens).
• Banners — errors, gentle warnings (e.g. “this model can’t call tools”), and a Resume banner if a run stalled.
• Star a message to pin it to the sidebar highlights for later.

The right-hand pane

• Tool runs — the full timeline of the worker’s tool activity, with token-saved estimates.
• 🌐 Browser — opens an in-app web browser the worker drives; you can watch it navigate, and replay the session with the scrubber.
• 🖥 Desktop — if computer control is enabled (off by default), a live screencast of the worker using your desktop.

The activity badge

Top-right, a small indicator (thinking / reading / typing / mining / web) tells you what the worker is doing right now, driven by its most recent tool.

Talk to Theodor first

Not sure who should do a job? Just ask Theodor. He sees your whole workspace and will either do it, delegate it to the right worker, or draft a Plan for bigger efforts.

Sending things out

If a request involves sending an email or posting publicly, the worker drafts it and pauses for your approval by default — it won’t send on its own. That’s intentional.

Tips

• Be concrete. “Summarize this PDF into 5 bullets and save to the Library” beats “look at this.”
• Queue while it works. Type your next instruction during a reply; it dispatches automatically.
• Watch the right pane to catch a worker heading the wrong way early — you can interrupt and correct.

In practice

Watching the right pane is the skill that separates good results from frustrating ones. A worker heading down the wrong path is visible in its tool runs long before it shows up in the reply — a search with the wrong terms, a file read from the wrong folder. Catch it there and one corrective message saves the whole run. Read only the final answer and you’ll be re-prompting from scratch.

Ask Theodor when you don’t know who should do the job. He sees the whole workspace, so “who should handle this, and can you get it started?” is a legitimate opening move. He’ll either do it, hand it to the right worker, or draft a Plan if it’s bigger than one task.

The task templates are filtered to what this worker can actually do. That makes the strip a reliable answer to “what is this worker for?” — if a template isn’t listed, this worker lacks the tool for it. Faster than opening the profile to read its toolset.

Queue your next instruction while it works. Typing during a reply is supported; the message dispatches when the current turn ends. Useful when you already know the follow-up and don’t want to sit watching the stream.

Outbound stays a draft by default. Ask for an email and you get an email you have to send. That is intentional, and worth keeping until a given workflow has earned your trust — see Budgets & approvals.

🎓 Learn it hands-on: Meet Theodor, your chief-of-staff

Terms in this page

• Worker / persona — a configured AI character you converse with; each has its own model, tools, and memory.
• Theodor — the built-in orchestrator (“chief-of-staff”) that coordinates other workers and drafts plans.
• Tool / tool call — a capability the worker invokes (web search, file read/write, browser, etc.). A “tool run” is one such invocation plus its result.
• Token — the unit AI models read/write in (~¾ of a word); providers bill per token, so the app shows token-saved estimates when results are reused.
• Cache (🔁) — a stored result from an identical earlier tool call, reused to avoid repeating work and spending tokens again.
• Full-duplex voice — a live voice mode where you and the worker can speak and be heard at the same time (like a phone call), rather than taking strict turns.
• Computer control / computer-use — an advanced, off-by-default capability that lets a worker take screenshots and operate your mouse/keyboard; every destructive action is approval-gated.
• Streaming — the reply appears word-by-word as the model generates it, instead of all at once.