> ## Documentation Index
> Fetch the complete documentation index at: https://www.rumus.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.
# Fleet Mode
> One sentence, safe ops across a group of saved hosts. Tell the AI agent what to do, review the command and its risk, then run it everywhere — with canary rollout and per-host results.
**Fleet Mode** runs an operations task across many [saved SSH hosts](/docs/remote/hosts) at once. Instead of opening a terminal per server or scripting a loop, you pick the target hosts, tell the AI agent in one sentence what to do, and review a single approval card before anything runs. The agent adapts the command to each host's operating system, and you get per-host output in one place.
> **One sentence, safe ops across a group of hosts.**
Fleet Mode is built around a simple promise: **every batch is reviewed before it runs.** Nothing executes on your servers until you've seen the exact command and its risk level and clicked **Allow**.
## How it works
Click the **+** in the tab bar and choose **Fleet Mode**. Fleet opens as its own full-page tab — not a terminal.
Click **Select targets** and choose from your encrypted [host vault](/docs/account/vault). Hosts are organized by your [host groups](/docs/remote/groups-known-hosts); selecting a group selects every host under it. Search by name or IP, or use **Select all** / **Clear**.
Type a plain-language instruction — for example, *"List disk usage across all hosts and highlight anything above 80%."* The agent reads your selected targets, detects each host's OS, and adapts the command if needed.
The agent proposes a batch as an **approval card** showing the exact command (one per OS), a risk level, and how many hosts it will touch. Choose **Allow**, **Canary**, or **Reject**.
A status HUD tracks every host live — connecting, running, succeeded, failed. Click any host to read its full output, copy it, or feed it back to the agent.
## Picking targets
Fleet targets come from the same encrypted vault as your SSH hosts — Fleet doesn't store separate credentials. The target picker shows your hosts grouped exactly as in [host management](/docs/remote/groups-known-hosts):
* **Groups** — select a whole group to select every host inside it.
* **Ungrouped** — hosts that aren't in any group live here.
* **Search** — filter the list by host name or IP.
* **Select all** / **Clear** — bulk toggles.
You must select **at least one** host before you can send a message.
**Targets lock once the conversation begins.** After you send your first message, the target list for that chat is fixed — this keeps the agent reasoning about a stable set of hosts. To run against a different set, click the **+** button to start a new Fleet chat.
## Telling the agent what to do
You write in natural language, not shell. The agent has two Fleet tools:
| Tool | What it does | Approval |
| ----------------------- | -------------------------------------------------------------- | :----------------------------: |
| **Check Fleet targets** | Reads the hosts you selected, including each one's detected OS | Read-only — runs automatically |
| **Run on Fleet hosts** | Proposes a command (one per OS) to run across the fleet | Always asks first |
Because hosts in a fleet may run different operating systems, the agent groups them into **OS buckets** (Linux, macOS, Windows) and writes one command per bucket. A single approval card can therefore contain a `df -h` for your Linux hosts and the equivalent for macOS — adapted automatically.
Some prompts to start from:
* *"List disk usage across all hosts and highlight anything above 80%."*
* *"Show the current load average and uptime for every host."*
* *"List the top 5 CPU-consuming processes on each host."*
## The approval card
No command touches your servers until you approve it. Each batch the agent proposes renders inline as a **Fleet batch approval** card:
* **The command(s)** — grouped by OS, e.g. *Linux hosts (8)* with its command.
* **Risk level** — a badge: **Low risk**, **Medium risk**, or **High risk**.
* **Scope** — *"Will run on N hosts."*
* **A warning banner** on high-risk batches: *"This action may affect multiple production hosts. Review carefully before approving."*
You have three choices:
| Action | What happens |
| ---------- | ---------------------------------------------------------------------------------- |
| **Allow** | Runs on every selected host. |
| **Canary** | Runs on a small subset first, then asks before rolling out to the rest. |
| **Reject** | Cancels the batch — nothing runs. The card shows *"Rejected — batch did not run."* |
### Risk levels
Rumus scores every proposed command against a set of patterns and labels it automatically:
* **High risk** — destructive or system-altering commands: `rm -rf`, `mkfs`, `dd if=`, `shutdown`, `reboot`, `systemctl stop` / `disable` / `mask`, `drop table` / `drop database`, `chmod -R 777`, writes to `/dev/sd*`, and fork bombs.
* **Medium risk** — commands containing `sudo`, `install`, `upgrade`, or `delete`.
* **Low risk** — everything else (read-only inspection, listing, status checks).
Risk scoring is pattern-based, not context-aware — it flags `rm -rf` even when the target is a harmless temp directory. Treat the badge as a prompt to look closely, and use **Canary** when you want a real-world check before committing to the whole fleet.
## Canary rollout
Canary runs the batch on a few hosts first so you can confirm it behaves before it hits everything. Rumus suggests Canary automatically for any **high-risk** batch, and for any **medium-risk** batch that targets five or more hosts.
1. On the approval card, use the stepper to set how many hosts go first, then click **Canary** (the button reads **Run on N first**).
2. The canary hosts run while the rest wait.
3. **If the canary succeeds**, you're prompted: *"Canary succeeded (N/total). Continue on the remaining N hosts?"* — click **Continue rollout** to finish, or abort to stop with partial success.
4. **If the canary fails**, rollout stops automatically: *"Canary failed (N hosts). Rollout stopped."* Nothing runs on the remaining hosts.
The default canary size is configurable — see [Settings](#settings) below.
## Watching execution
While a batch runs, a status HUD sits above the chat input with a live summary: **"3 done · 1 failed · 2 running / 6 total."** Each host moves through these states:
| State | Meaning |
| --------------- | ----------------------------------- |
| **Queued** | Waiting to start. |
| **Connecting…** | Opening the SSH connection. |
| **Probing…** | Detecting the host's OS. |
| **Connected** | SSH ready, command not yet run. |
| **Running** | Command executing. |
| **Succeeded** | Finished with exit code 0. |
| **Failed** | Finished with a non-zero exit code. |
| **Aborted** | Stopped by you mid-run. |
From the HUD you can:
* **Abort** — stop the running batch immediately.
* **Retry N** — re-run only the hosts that failed.
* **Log** — open the per-host output viewer.
Hosts in a batch run in parallel, so a large fleet finishes in roughly the time the slowest host takes.
## Reading per-host output
Click any host (in the HUD or the approval card) to open its **Execution log**. Each log shows that host's `stdout` and `stderr` (errors in red), with these actions:
* **Add to Agent** — select any text in the log and send it back to the agent, with the host name attached, to ask a follow-up.
* **Download** — save the output as a `.log` or `.txt` file.
* **Copy** — copy the full output to the clipboard.
## Settings
Fleet preferences live under **Settings → Fleet**:
| Setting | What it controls | Default |
| ------------------------- | ----------------------------------------------------------------------------------- | :---------: |
| **Canary execution mode** | Whether the canary batch is a **fixed count** of hosts or a **ratio** (percentage). | Fixed count |
| **Canary host count** | How many hosts run first, when mode is *fixed count*. | 1 |
| **Canary ratio (%)** | What percentage of hosts run first, when mode is *ratio*. | 20% |
| **Auto probe OS** | Detect each host's OS before running, so the agent can adapt the command. | On |
| **Connect timeout (s)** | Mark a host as failed if its SSH handshake takes longer. Range 5–300. | 20 |
## Fleet vs. multi-terminal broadcast
Both run something across several hosts, but they're for different jobs:
* **[Multi-terminal broadcast](/docs/terminal/canvas)** types the same keystrokes into several live terminals at once. You drive it directly, in real time, with full TTY control.
* **Fleet Mode** is AI-assisted and batch-oriented: you describe the goal in words, the agent writes (and OS-adapts) the command, and you approve it once with a risk review and canary rollout.
Reach for broadcast when you want hands-on, interactive control; reach for Fleet when you want a reviewed, one-shot operation across a saved group of hosts.
## History
Fleet conversations are saved like any other chat. Reopen one from your history and it loads with the same targets, messages, and approval cards — a record of exactly what was approved and how each host responded.
## Troubleshooting
Fleet uses your saved known-hosts entries. If a host key isn't trusted yet, the host fails with *"Host key not trusted — try again to approve, or check the host's known-hosts entry."* Connect to that host once in a normal [SSH terminal](/docs/remote/hosts) to accept its key, then retry the batch. See [Known hosts](/docs/remote/groups-known-hosts).
The SSH handshake didn't finish within the connect timeout. Confirm the host is reachable and that it works from a normal terminal, then raise **Settings → Fleet → Connect timeout** for slow or distant hosts.
That's expected when your hosts run different operating systems — the agent writes one command per OS bucket so each host gets a command that actually works there. Review each bucket on the approval card.
Open each failed host's **Execution log** to see its `stderr`, fix the cause, and use **Retry** in the HUD to re-run only the failed hosts.
Targets lock once a conversation starts. Click the **+** button to begin a new Fleet chat with a different set of hosts.
Fleet question we didn't cover? Ask in the [Rumus community](https://www.rumus.ai/community).
## Next steps
Save the hosts Fleet will operate on.
Organize hosts into groups you can select as a fleet in one click.
How the AI agent proposes and runs commands.
Interactive, real-time fan-out across live terminals.