SAM Assistant (preview)

The SAM assistant is one private conversation for your whole account. Ask it what needs your attention or what’s running across your projects, and have it start work in a project, reply in a project chat, or stop work. Anything that touches a repository still happens in an ordinary project chat you can open. It’s a preview, and it’s off until you turn it on.

Turn it on

1. Open Settings. The SAM assistant card sits at the top, above the tabs.
2. Turn the switch to On.
3. Select Open SAM →. SAM also appears in the main navigation, right after Home, while it’s on.

Turning it off stops new actions and keeps your history and any work it started. If the card says “SAM is temporarily unavailable on this installation.”, an administrator has turned it off for everyone.

What you can ask

An empty conversation offers a few starters: What needs my attention?, What is running across my projects? and Show my recent ideas. Selecting one only fills in the message box. Beyond those, you can ask it to:

• read your projects, your inbox, chats (list, read, search), tasks, ideas and agent profiles;
• start work in a project, or reply in a project chat, which it does straight away;
• file or update an idea;
• stop work or answer an agent that’s waiting, which it does only after you confirm.

Name the project, and say which profile, branch or review you want, as you would in a project chat. Work it starts is recorded as started by you. It runs where the profile’s runtime says. With Auto, or no profile, it runs on a VM only when you or the project has a cloud account connected, and otherwise in an Instant container, which has no devcontainer and no Docker. For builds and tests, name a Cloud VM profile.

It can’t change settings, connections, compute pools, automation or deployments, and it can’t read schedules, channels or a team of agents, or interrupt an agent. Use the app for those. Your AI app, through the Connector, can also read schedules, channels and teams, schedule work, and interrupt an agent. Requests that need repository files read, searched or changed always become a project chat or task: follow its Open chat link to see the actual work.

Watch it answer

While SAM works, the reply shows Starting, Thinking, Working or Writing, with the time elapsed, and each check as a line that updates when it finishes, such as Looking through your 20 most recent tasks in Website redesign… becoming Read 20 recent tasks in Website redesign. The answer then appears as it’s written. You can leave the page and come back.

SAM links only to projects, tasks, chats and ideas it actually looked up while answering. For questions about what’s running or what shipped, it reads live data and the reply lists what it checked, such as Checked your inbox. If an answer about current work lists nothing, ask it to check again.

Stop replaces Send while it works. Stopping ends SAM’s turn (“SAM · cancelled”) but not project work it already started.

What the cards under an answer mean

Each action SAM took, or wants to take, gets a card with a link to its target (Open chat, Open task, Open idea or Open project) and Exact request, the input SAM recorded. On a card that waits for you, Exact request is already open, so you can check what you’re confirming.

Card says	Meaning
Done	It happened
Needs your confirmation	Stopping work or answering an agent waits for you: Confirm stop or Keep running, Send answer or Don’t send
In progress	It’s still happening
Outcome unknown	SAM couldn’t confirm whether it happened. For new work, Check outcome looks for it and never starts it twice; for anything else, check the project before asking again
Declined	You chose not to
Not run	It didn’t run
No longer available	Its target is gone, or you can no longer see it

An agent form with values, such as a secret, has to be filled in the project chat itself: don’t paste secrets into the SAM conversation.

Project work, above the conversation, follows the work SAM started for 24 hours, and counts items that need attention or have new updates. Expand it for a card per task with Open chat. Mark seen clears the “new” marker; it doesn’t answer anything. After 24 hours the card says “Automatic updates expired. Open project work for current status.”

When SAM stops before answering

A reply marked SAM · interrupted reads “SAM could not finish this turn. Review recorded actions before continuing; uncertain actions are not automatically repeated.” It happens when:

• your AI usage budget ran out. Every answer counts toward your daily token limit and monthly cost cap. A short status question uses about 30,000 to 40,000 input tokens, so with the default limit of 500,000 input tokens a day you can run out after a dozen or so questions. See and change your limits in Settings → Usage, under Budget Controls;
• the turn took longer than 2 minutes or 20 steps;
• SAM couldn’t back its answer with data it read; or
• the AI model behind SAM failed or gave an empty answer.

Check the cards for anything it already did, then ask again.

Two limits show raw messages for now: more than 30 messages a minute shows “RATE_LIMITED” (wait a minute and send again), and a message over 16,000 characters shows “A bounded message and requestKey are required” (shorten it).

If a message doesn’t get through, it stays in the box, and Retry sends the same request again, even after a refresh, so it’s never sent twice.

History and privacy

The conversation opens on its most recent messages; Load older messages fetches more. Earlier SAM history, when present, holds messages from the earlier assistant, read-only.

The SAM conversation is private to you. Project chats and tasks it starts are visible to that project’s members, as usual, and SAM is told to pass them only what the request needs.

For self-hosters

The assistant arrived in v2026.10.21, and live answers in v2026.10.22. Each person turns it on for themselves. A superadmin can turn it off for everyone by setting operations.samOperatorDisabled to Enabled in Admin → Settings (the setting is named for turning SAM off). It runs on the SAM_MODEL model (GLM 5.3 Flash on Workers AI unless you change it, to a model from SAM’s platform model list), billed to your Cloudflare account. See configuration and the API contract.