Compass Chat

Click Chat in the top bar, from anywhere in Compass. Chat opens docked on the right, over the page you're on; drag its left edge to make it wider, and press Esc or click outside it to close it. Expand to full width gives the conversation the whole page, and Dock to the side brings it back.

You can also start from either Home page: type a question into its composer and chat opens full-width with your question already sent.

Chat is the same everywhere, and the model you pick — not the page you're on — decides which project a turn runs in. Closing chat keeps your place, including a half-written message.

A new chat opens with What should we work on? and the composer in the middle. Your chats are listed under Show chats — pinned to the side when chat is expanded, sliding over the conversation when it's docked — grouped by date and searchable. Pick one to carry on where you left off. New chat starts fresh, and Delete chat (on a chat's row, or in the title's menu) removes it straight away.

Chats belong to you: you see your own, across the organization. Deleting a chat removes its attached files, but the executions it produced stay in Observability.

Type your message and press Enter to send (Shift+Enter for a new line). Everything else sits on the composer itself:

• + — attach files.
• Tools — what chat can call; see below.
• The model — which model answers, and how hard it reasons.
• Model settings — temperature, top P, and max tokens.

The line beneath shows the connection in use, the Harness, and Bills to Project — the project the turn runs and is billed in.

Attachments

Attach files with +, by dragging them onto the composer, or by pasting. Chat takes images (PNG, JPEG, GIF, WebP), PDFs, and text files (plain text, Markdown, CSV, JSON), up to 20 MB each. Attachments reach whatever chat runs, including agents and workflows called as tools, and files from earlier turns stay available for the rest of the conversation. See Files & attachments.

The model chip opens a searchable list of model connections, grouped by project — it lists models from every project you belong to, though a turn only runs on a model from a project where you're a Developer or Admin. The turn runs, and is billed, in the project that owns the connection you pick, which is what Bills to Project shows. Chat remembers the model you used last.

For a model that can reason, the list's footer has a Reasoning control: Off (the default) or an effort level. More effort means more considered answers, more tokens, and slower replies. Models that always reason show Always on.

Model settings holds Temperature, Top P, and Max tokens. Settings a model doesn't accept are greyed out, and some are fixed while reasoning is on.

Tools chooses what chat may call. It lists four groups, all from the project chat is billing to:

• Agents and Workflows — tick one to add it. Each runs a version you choose: Live · vN runs the deployed version, Draft runs the latest edits. Deployed work defaults to Live. The shield on a picked row makes each call wait for your approval.
• Integrations — any integration that can be used as a tool (everything but Qdrant) with an active connection. Click one to pick which Actions chat may use, which of them wait for your approval, and which Connection, if there's more than one.
• Datasets — the project's datasets. Click one to pick its actions and approvals.

A newly added integration or dataset starts with only its read actions on — actions that change things are one tick away. Your selection applies to every chat billed to that project, and is remembered in this browser.

When you open chat on the workflow canvas with a model from the workflow's project, the workflow you're editing is added automatically, marked canvas, and runs its draft — the quickest way to try a change.

The Harness menu, beneath the composer, sets what chat brings to the conversation beyond the tools you picked:

• Compass (the default) — "For working with Compass: your workflows, agents, connections, and runs."
• None — "Basic chat. The model runs as configured, with nothing added."

What chat can do on the Compass harness

On the Compass harness, chat can look into Compass as you — with exactly your permissions, in any project you belong to:

• Your work — list projects, workflows, and agents; read a workflow's or agent's definition; see which are live and which have undeployed changes.
• Executions — find executions by name, workflow, or status (including those waiting for approval), and read a trace step by step to explain what happened or why something failed.
• The catalog — components, integrations and their actions, your connections (never their credentials), and model providers with what each model supports.
• Datasets — list them, see their documents, and search them.

It can also draft changes: create a new workflow or agent, or edit one's draft. Every change pauses for your approval first, and it only ever touches drafts — chat can't deploy. You review the draft and deploy it yourself. It can't delete anything, create connections, or decide approvals.

The Compass harness also keeps long conversations working: it finds its Compass tools on demand rather than loading them all up front, reads very large results a slice at a time, and summarizes older history once a conversation grows long.

A reply streams in as it's written. Along the way you'll see:

• Reasoning — when reasoning is on, the model's thinking, collapsed once the answer starts.
• Tool calls — each as its own row, with a version chip for agents and workflows (v3 · Live, Draft), how long it took, and whether it succeeded. Click a row for its Input and Output.
• Approval requests — when something needs sign-off, a card under the tool call shows its exact inputs: Yes, run it, or No, reject and tell the agent why, with an optional reason. Once decided, the card folds into a receipt the call keeps — who decided and when. See Human-in-the-loop approvals.

Under each finished reply: Copy reply, the turn's tokens and cost — including any agents and workflows it called — and Trace, which opens the full execution tree: every step, tool call, and child execution, with inputs and outputs. It's the same trace you get in Observability.

Each turn is an execution in the project that owns its model, listed in Observability as Compass Chat Agent and tagged System. Agents and workflows it calls appear as child executions. Project and organization spend limits apply to chat like anything else.