Workspace

If you are already signed in to Trinity, open Workspace from the nav — it opens in its own browser tab, your platform session is your Workspace session, and your roster includes both the agents you own and the ones shared with you.

An external client signs in with email instead:

Codes expire after a few minutes. Resend code becomes available after a short cooldown, and ← different emaillets you switch before verifying. Signing in with an email that has no agents shared with it still works — you land on an empty roster.

A client session slides: it renews while you use the Workspace, ends after the idle window, and never outlives the cap (defaults: 7 idle days, 30 days total). When it times out, the sign-in form says so — Your session timed out — and signing in again returns you to where you were. An administrator sees the policy under Settings → Retention → Workspace sessions; changing it needs an entitlement.

The button at the bottom of the sidebar signs out. For a platform user it is labelled Sign out of Trinity, because your Workspace session is your platform session and ending one ends the other. A client's expired session never turns into the operator's: on a browser that also holds an operator login, the form offers Continue as user@example.com as an explicit click rather than switching identity on its own.

The main app and every Workspace tab in the same browser share one platform sign-in, and open tabs follow it. Sign out and back in on one tab, and a Workspace tab left open from the old session picks up the new one instead of ending it. Sign out on one tab, and the other tabs drop the session too, without a background tab jumping to the sign-in page. A client signed in with an email code is not sent to the operator sign-in page just because the same browser also holds an expired operator login.

Three columns on a desktop screen: the sidebar on the left, the conversation in the middle, and a side panel on the right (the rail, or the agent's canvas during a voice call). The seam beside the sidebar, and the one beside the rail while it is open, are drag handles — drag to resize, double-click to reset. The conversation takes whatever is left, and the panel's maximum follows the viewport, so a width set on a large screen is clamped on a smaller one. On a phone the sidebar becomes a drawer behind the Menu button, and the rail becomes a strip above the composer that opens a bottom sheet.

Theme. The right end of every chat and room header carries the theme switch. It names what is on screen — Light, Dark, or System · dark / System · light when the choice follows your operating system — and on a phone it shows only the icon. Click it and pick Light, Dark or System (arrow keys move between them; Esc closes the menu without touching a running turn or call). It is the same setting as the theme control in the main app, remembered in this browser, and switching keeps your place, your draft and your thread. External clients get it too.

Top to bottom:

Every chat you have with the active agent is a tab above the thread. Main is always first; the rest follow most recent first, and the ones that don't fit collect under N more. Press New chat and a provisional New chat tab appears at once; the chat itself is created when you send the first message.

Main is named by its role and cannot be renamed. Reset, in the header on Main only, archives the current Main and starts the agent cold — no confirmation, because nothing is lost: the archived chat stays in your list as an ordinary chat (named and dated if it had no title), a line in the new Main names it, and it becomes the newest tab. Reset is refused while a reply is in flight, and resetting a Main you have never used does nothing.

Every other chat can be renamed in place — from the pencil beside its title in the header, or on its sidebar row. Enter or clicking away saves, Esc abandons. A title is one line of up to 100 characters. Otherwise Trinity names a chat for you from your first message, and takes one more pass on the second exchange if the first was only a greeting. A name you typed always stands.

One box: the message field on top, the controls in a row inside it. Enter sends; Shift+Enter adds a line.

An agent that is stopped or unreachable is labelled above the field before you type; the message still sends, and the server's refusal is the answer.

Replies stream as they happen. Under your message a live card shows the agent's status, the elapsed time and one line saying what the agent is doing right now, with Stop and Open in Work — the same card the rail's Work tab lists afterwards. When you send, the chat scrolls down to the card, unless you had scrolled up to read something.

The activity line reads the same way everywhere: Reading .../src/app.py, Running pytest -q, Searching for "pattern", Fetching example.com, Using github, Delegating to researcher, Writing a reply, or Thinkingbetween steps. A new line slides up over the old one and stays long enough to read; a burst of quick steps shows only the latest. On a quiet stretch the last line stays, and the line clears when the run ends. Your own turn's line comes from its live stream. Every other live run — a delegated job, a scheduled run, a room turn — takes its line from the agent's regular heartbeat, which a platform user sees on the Work tab's cards and on a room's cards. A heartbeat line older than 30 seconds is dropped rather than shown as current, and a card with nothing to report leaves the line empty rather than guessing.

Stop, or Escwith nothing else open, cancels the in-flight turn and puts your words back in the composer (in front of anything you typed while waiting). A cancelled turn shows as cancelled, not as an error. Esc does nothing when no turn is running, and yields to whatever is open on top — the typeahead, the agent picker, dictation, a file preview.

Turns on one chat are serialized to protect the agent's memory, so a second message while one is running is refused with This conversation is already handling a message — wait, or start another chat for parallel work. If a send fails you get the real reason (busy, timed out, too large, too many messages, the agent's usage limit) instead of a bare failure, and Retryis offered only when nothing reached the agent — a turn the agent already ran is not silently billed twice.

Closing the tab doesn't stop anything: the turn keeps running on the server and the reply is waiting when you come back.

The rail sits beside the conversation, collapsed to a strip of icons by default; click one to open it, Collapseto close it. Its state — open or collapsed, and which tab — is remembered across chats and reloads. A dot on a tab means something happened since you last looked: a pulsing dot for work or a loop in progress, a plain dot for a canvas or file updated since you last opened that tab.

Files. Drop a file on the tab, or click to send one. In a room, Send to defaults to Everyone in this chat (N agents) — one copy per agent, the same as dropping the file on the room itself — and still lists each agent for a single recipient. The receipt names who got it (Sent “shot.png” to analyst and sidekick.). A file counts as sent only when every recipient got it; otherwise the error line names the file and the agents it missed. The list is grouped Files you sent / Files from {agent}. Click a name to preview it — images, Markdown and text up to 256 KB; ← and → step through the previewable files, Esc closes — or Download to save it. The bin icon deletes a file you sent, or removes a shared file from your list without touching the share. An agent's owner, signed in as a platform user, additionally gets Delete for everyone.

Canvas. One canvas shows at a time; pick another from the chips above it, where a pinned canvas carries 📌 and comes first. Once an agent has more than six, Search canvases… filters them by title or id. The header states two facts and draws no conclusion from them — Updated 2h ago · agent last ran 40m ago. PDF prints the open canvas through the browser's own print dialog. An agent's owner (or an admin), signed in as a platform user, also gets Manage: each row shows its age, a pin toggle and Delete, and a checkbox per row feeds Delete selected with one confirmation naming the count; everyone else has a read-only panel. Sharing a canvas at a link is done from the agent's Canvas tab on Agent Detail, not from the rail — see Agent Canvas. The canvas you have open travels with each message you send from that chat, so add a column to this names the right one. It is context, not permission: the agent still reaches only the canvases it could already reach, and a canvas deleted mid-conversation resolves to nothing.

Clicking an agent no longer opens a report about it. Its numbers sit in a band under the header of every chat with it — tasks in the last 7 days, the share completed, the share that succeeded first try, a Helpful / Not helpful tally where people have rated it, and a small activity chart. Everything else is the rail's Info tab:

It reports; it does not configure. There are no schedules, skills, logs, costs or model details here. A question the agent raised appears above the composer of the chat it belongs to, and answering it there tells you whether the agent is picking the work up — see Approvals. Everything is read from stored data, so a stopped agent still renders. The old address /workspace/a/{agent} still works and lands in the chat.

The microphone in the composer (Speak your message) turns speech into text in the message field. Nothing sends until you press Enter. Clients and platform users both get it, and it uses one of two engines:

When neither engine can work, the composer shows no microphone at all rather than one that fails on every press. ElevenLabs grants permissions per endpoint, so a key that speaks replies aloud may still lack Speech to Text. That turns platform transcription off; the browser's engine still works where there is one.

A failed transcription gets a line above the message field that says why: the key lacks the speech-to-text permission, the key was rejected, the account is out of credits, too many voice messages in a short time, the recording could not be read, or the provider failed. You can always type instead.

For admins. Settings → General → Voice (ElevenLabs) shows, beside the key's configured badge, whether the key can transcribe:

Trinity checks each key once and repeats the check every few hours, and at once when you save the key again. A transcription the provider refuses (HTTP 401 or 403) also marks the key, so the next page load stops offering platform transcription. Under the badge, Last voice-input failurenames the most recent failed transcription: the cause, the provider's HTTP status and status word, and the time. It stays for 24 hours, or until you save the key again. The key itself is never shown.

Mention another agent from a 1:1 — type @ and pick it — and Workspace opens a room containing both and posts your message there; the original 1:1 is left as it was.

Files you attached in that 1:1 and have not yet sent with a message go along: the composer's attachment chips, and files sent from the rail's Files tab in the last 15 minutes. Workspace waits for uploads still in progress, then delivers each file to the agents that do not have it yet, before posting your message, so the mentioned agent can see what you asked about. A line under the room's composer then says what happened — Sent with your message: shot.png — also delivered to sidekick., shot.png didn't reach sidekick — attach it again here to retry., or that a file was not carried over because it never finished uploading — until you dismiss it or send your next message. If the room cannot be opened, your text comes back to the 1:1 composer with its attachment chips.

Inside a room, + Add agent recruits another, and only a person can do that. Rooms carry participant avatars, a star and a budget warning as the conversation approaches its cap. For a platform user, each agent working on the room's message gets its own live card with Stop. A room's cards, files, names and rules are covered in Shared Sessions (Rooms). Against an older backend without rooms, the picker is single-select and an @name stays ordinary text.

Workspace surfaces the agents already shared with a person's email — there is no separate access list. To give a client agents, share each one with their email using the normal agent sharing and access control model; remove the share to revoke access.

Capability cards come from the agent's exposed playbooks where an operator has configured them, and otherwise from the use_cases in its template — so what a client sees you can shape without touching Workspace itself. The model an external client's turns run on is the one set for the agent's public channels on its Sharing tab, falling back to the platform default.

Workspace is a client-facing shell over the platform's existing agent behavior, not a new agent capability, and no MCP tool targets it — an agent reaches the people in it through its replies, its reports, its questions and its canvas. It is served by /api/enterprise/client-portal/*(the prefix is historical), authenticated by either a Workspace session token or a platform JWT, with every per-agent route scoped to the caller's roster. Agent-scoped MCP keys are refused.

Asks and the Work tab have their own routes under the same prefix — see Approvals and Executions.

Full request/response schemas are in the backend's Swagger docs (http://localhost:8000/docs when running locally).