Platform

Chat

The composer is the primary surface. Every message you send flows through a single turn pipeline that picks the right model for your tier, exposes the right tools to the agent, narrates work in flight, and surfaces results in chat plus the right side panel. This page covers everything you can do from the composer.

Updated today

Overview

At a glance

Position: Bottom-centre of the screen
Submit: Enter to send, Shift Enter for a newline
Slash key: Opens the Workforce picker over the composer
Plus button: File attachments, screenshots, links
Tier picker: Bottom-right of the composer; Minuet, Allegro, Forte
Donut: Spins when long-running work is in flight
Reply streaming: Token-by-token, with prose narration between tool rounds
Quick replies: Suggested follow-ups appear under the agent reply

The composer is intentionally minimal. Everything that affects the turn is reachable from it: the tier picker, the Workforce picker, the attachment menu, the submit action. The chrome you see around it (the right panel tabs, the donut, the Brain summaries) reacts to what the agent is doing in real time but never takes input.

The composer

The composer grows vertically as you type. Long pastes do not blow out the layout. The Submit button activates as soon as there is text or an attachment. Empty submissions are ignored.

Action	How
Send the message	Enter
Newline without sending	Shift Enter
Open the Workforce picker	Forward slash at the start of a message
Attach a file	Click the plus button to the left of the composer
Cancel the turn	Stop button replaces submit while a turn is running
Switch tier	Click the dots picker bottom-right
Open Background Jobs panel	Click the donut under the composer

Tier picker

Three levels, persists across the conversation, applies to every part of the turn.

The picker is the source of truth for which model class runs the turn. Slash commands, canvas generation, and skill runs all honour it. Free accounts can only pick Minuet; paid accounts can pick any of the three.

Tier	What it is	When to pick it
`Minuet`	Cheapest tier. One dot indicator.	Everyday conversations, quick lookups, summarising attachments, writing single-paragraph drafts, exploring an idea.
`Allegro`	Default for paid accounts. Two dots indicator.	Multi-step skills, full canvases, complex outreach sequences, research tasks that need to plan across several rounds.
`Forte`	Highest reasoning quality. Three dots indicator.	Strategic synthesis, multi-document analysis, hard debugging, decisions that hinge on subtle judgement.

Tip

Pick once at the start of the turn. The picker remembers your choice across turns in the same conversation and applies it to the next message automatically.

Slash commands and the Workforce picker

Press the forward slash key at the start of an empty composer (or right after a space) to open the Workforce picker. The picker is searchable and groups by agent. Pick an agent to scope your message to that agent. Pick a specific skill to fill the composer with the skill prompt template ready for you to add your context.

01
Open the picker
Forward slash in the composer. The picker opens above the input.
02
Pick an agent or a skill
Type to filter. Up and down arrows move the cursor. Enter selects. Examples: /cortex for research, /specter for outreach, /sentinel for engineering tasks.
03
Add the rest of the message
The picker closes. The slash command stays as a tag at the start of the composer. Type the rest of your message normally and submit.

Note

You do not have to use a slash command. A plain English message works for almost every case; the agent resolves to a skill automatically when one matches.

Attachments

Anything you can paste, drag, or pick from your file system.

File type	What the agent sees
Images (PNG, JPG, WEBP, SVG)	Visual analysis on the fly. Useful for screenshots, mockups, photos.
PDFs	Full text extracted and provided as context. Long PDFs are summarised into a working buffer.
Text and markdown	Inlined verbatim. Up to ~50K characters before truncation.
Audio (MP3, WAV, M4A)	Transcribed automatically. The transcript becomes the agent's input.
Office documents (DOCX, XLSX)	Converted to text. Spreadsheets are flattened into a table view.
URLs in the message	The agent decides whether to fetch. Plain links are kept as references.

Narration and tool rounds

What you see in real time while the agent works.

Every turn that uses tools interleaves prose narration with tool calls. The pattern is consistent: one short sentence describing what is about to happen, then the tools fire, then a short observation about what came back, then the next set of tools, and so on. The final summary comes last.

Note

Tool blocks are collapsed by default. Click any tool row to expand the raw input and output. The same view is used across every tool category: shell, skills, web fetches, integrations, canvas generation.

Quick replies

Suggested follow-ups attached to an agent reply.

When an agent finishes a turn it sometimes attaches a row of chips with suggested next messages. Click any chip to send that text as your next message. The chips are scoped to the immediate context of the reply (yes or no, approve a plan, pick a variation, save a draft).

Quick replies are skippable. Ignoring them and typing your own message works identically.

Memory injection

What the agent already knows before you send a message.

Before every turn the platform pulls the most relevant items from your Brain and adds them as context for the agent. Profile, key memories, recently saved facts, and any open work in flight. You do not have to remind the agent of context you already gave it earlier in the conversation or in past sessions.

The retrieval is hybrid: semantic similarity for prose memories, tag and recency filters for working notes, and structured lookups for profile fields. The agent receives a short stitched block, not a wall of raw text.

Sessions and the share surface

One chat is a session. Sessions persist, are searchable, and can be shared.

At a glance

Session lifetime: Until you delete or archive it. Conversations live in the left sidebar.
Search: Search across all sessions from the home tab or the sidebar.
Share: Each session can be turned into a read-only share link from the session menu.
Forking: Branching a session creates a new conversation pre-loaded with the prior context.
Archive: Archived sessions stop appearing in the active list but stay searchable.

Note

Shared sessions render with read-only chrome and skip the right-panel tools. Anyone with the link can read; nobody can resume the conversation without an account.