Docs

1. Install the desktop app

Go to the download page to get the Windows / macOS / Linux client and install it. After launch there are two permanent entry points at the top: Home (assets and sessions) and Deck (the ops assistant). Everyday assets live in the Home sidebar: hosts, keys, forwarding, jobs and settings.

• In Settings you can change the default local terminal (the assistant and local shell run there) and the system language.
• You can register devices and open sessions without signing in; sign in when you want cloud models and subscription features.

2. Sign in — cloud models work immediately

The assistant, command suggestions and monitoring exploration use cloud models by default: sign in and go, no model keys to configure.

1. On the sign-in page sign in with email or Google / GitHub, then choose "Authorise this machine"; the desktop app picks up the current account.
2. The cloud model list is delivered with your account and can be picked directly from the model selector in the Deck input bar.
3. Usage counts against your subscription quota: the monthly quota is used first, then your balance; the trial period includes cloud quota too.
4. Check the current status, monthly quota usage and balance under Account → Subscription.

Need more quota and priority support? Upgrade to Pro inside the desktop app; renewals are managed by the app store. See Pricing.

3. Register devices and keys

Sessions, the assistant, file transfer and forwarding share one set of assets. Put "who to connect to and with what" into the inventory first, so you never have to piece connection details together again.

Create a key (recommended first)

1. Home → Keys → New key.
2. Paste the public key and private key; add the passphrase if there is one; paste the certificate if you need one.
3. Only the OpenSSH format is supported; convert PuTTY PPK files to OpenSSH first.

Add a device

1. Home → Hosts → Add device (the empty state also prompts "No devices yet, click to add").
2. Basic info: name, host, type and group. The type affects command suggestions and some device-side interaction, so choose it carefully.
3. Fill in SSH (port 22 by default, username, password or key), Telnet (port 23 by default) and/or RDP (domain, resolution, security layer, etc.) as needed.
4. SSH has extra options for session reuse, SSH agent and auth method; Telnet has prompt detection and command echo verification; RDP has colour depth, keyboard layout and drive redirection.
5. Save, or Save and connect straight away.

Passwords, keys and other sensitive assets stay in local data; signing in to the website never uploads them as a cloud inventory.

4. Open a session

Confirm the target is reachable before handing it to the assistant. Click a host card and choose the protocol to open.

• SSH / Telnet: opens a terminal pane. The first connection asks you to confirm the host fingerprint; a changed fingerprint raises an alert and you can reject, trust once or trust permanently.
• SFTP: uses the same identity as SSH and opens the dual-pane file view.
• RDP: enters the Windows desktop with clipboard and local drives shared.
• Local shell: pick a terminal from the hosts toolbar (Command Prompt / PowerShell / WSL, etc., depending on what is found on your machine).
• After a disconnect use Reconnect; the + at the top opens a new tab to revisit recent connections.
• Adjust terminal settings (font, theme) and find inside the pane without interrupting the session.

Tabs can be dragged to reorder or merged into split views: merge when you need two devices side by side, drag back to the tab bar when you need to focus.

5. File transfer and tunnels

Dual-pane files

Open SFTP inside a session to see local and remote side by side. Common actions: open, copy to the target directory, new folder/file, rename, delete, change permissions, show hidden files, copy path. On conflicts choose skip or overwrite as prompted; progress stays at the bottom of the pane.

Tunnels and forwarding

1. Home → Forwarding → New forward.
2. Choose a type: Local (a local port into the target environment), Remote (a remote port back to this side) or Dynamic (SOCKS5 proxy).
3. Fill in name, port, bind address, target host and address, then use Start forwarding to toggle it. States are idle, running or error.

Files and forwarding do not depend on the assistant and can be used on their own.

6. Use the ops assistant (Deck)

Switch to Deck at the top. The left side holds new chat, skills, MCP servers, recent chats and device references; the middle is the conversation and tool execution area.

Your first conversation

1. Make sure a model is selected in the toolbar; cloud models are available right after sign-in.
2. Keep the approval mode at Ask for approval (the default): you click Allow or Skip for every action.
3. Use the + on the left of the input bar to add devices, groups, skills or images, or drag from Device references. References become context chips so the assistant knows the scope.
4. Describe the goal in natural language. Start with a read-only inspection, such as CPU, disk and key interface status.
5. When an action awaits confirmation choose Allow or Skip; for repeated similar commands use "Remember allow" to avoid being interrupted for the same pattern.

Approval modes

• Ask for approval: you click Allow or Skip for every action.
• Audit for me: approval is only requested for actions detected as risky.
• Allow all: no per-step prompts, but dangerous commands are still blocked; the danger word list is customisable. Do not use this level until you know the flow.

7. Built-in tools

The assistant ships with 39 built-in tools covering device execution, files and documents, asset lookup, code search, web access and monitoring. Changing actions go through the approval mode first; queries run directly.

Device execution

• Local command: runs commands on this machine; long tasks move to the background and you can list processes, read output or terminate at any time.
• SSH / batch SSH: runs commands on one or many registered devices, concurrently in batch mode.
• Telnet: runs commands on Telnet targets such as network devices.

Files and documents

• SFTP transfer: upload and download between this machine and devices.
• Read / edit files: read local and remote text files and apply exact replacements.
• Document processing: read PDFs, read and write Excel and Word, and produce inspection reports directly.

Assets and context

• Find devices / groups / skills: look up inventory assets as context at any time.
• Chat history and task list: browse past conversations and keep multi-step work as a checklist.

Code and references

• Code tools: repository structure map, content search, symbol definitions and references.
• Explore sub-agent: read-only parallel exploration of large directories or long documents, summarised back into the main chat.
• Web search / fetch page: search external references and fetch page content; web search comes with your account.

Monitoring and skills

• Metric read/write: read device monitoring metrics and, after confirmation, add or modify metrics.
• Skill distillation: write the current procedure as a new skill or update an existing one (requires confirmation).

Typical scenarios

• Inspection: check CPU, disk and key interface status across devices and produce a readable conclusion.
• Change: edit a config file and restart the service, confirming step by step.
• Transfer: package remote logs, pull them back, analyse and output a report.
• Troubleshooting: combine web search and document reading to locate the source of an error.

8. Skills

A skill is a reusable Markdown prompt (name + one-line description + body), stored in the local skill library and included in cloud backup.

• Create: write one under Deck → Skills; distil one from past conversations with AI; or let the assistant save a new skill after your confirmation.
• Use: "+ → Skills" in the input bar, or reference it directly. Once referenced it stays active for the turn, so follow-ups need no repeat reference.
• Evolve: the assistant can refine skill content while using it; changes also require confirmation.

9. MCP servers

The assistant can connect to MCP (Model Context Protocol) servers to bring external tools into the same conversation.

• Transports: stdio (launched by a local command) and Streamable HTTP (remote URL).
• Configure: Deck → MCP servers → Add; fill in command, arguments and environment variables, or paste a JSON config to import.
• Activation: once enabled, its tools join the assistant's tool set; built-in tools win on name clashes. The list shows each server's status and tool count.
• Safety: every MCP tool call goes through the same approval mode as built-in tools.

Calls do nothing while a server is not ready (process not started or URL unreachable); check its status in the list first.

10. Terminal command suggestions

Turn on Command suggestions under Settings → AI capabilities and candidates appear next to the cursor while typing in local or remote terminals. The suggestion model is configured in the cloud; locally you only choose the candidate count.

• Preview and list side by side; Tab to accept, arrow keys to switch; close the panel to keep typing.
• The panel accepts a natural-language intent: press Enter to generate candidates.
• Candidates follow the current session and device type, matching the dialect of the moment.
• If suggestions do not appear: check the toggle is on, you are signed in, and the current pane is a session type that supports suggestions.

11. Monitoring and discovery

Path: Home → Jobs. Two tabs: Monitor and Discover.

Monitor

1. Click New task, tick the hosts to observe and confirm.
2. View metrics on the device card; open details to configure name, protocol, interval, parsing, alerts and AI escalation.
3. AI escalation options: off, analysis only (L1), repair with confirmation (L2). L2 still goes through assistant confirmation and never changes config silently.
4. Use "Collect now" to verify a single metric produces the expected value.

Discover

1. Choose an exploration model under Settings → AI capabilities (cloud models are available after sign-in).
2. Back in Monitor, start Auto-discover metrics on a device; the process shows as exploring.
3. Review the exploration summary and candidates on the Discover tab; each is pending or adopted. Tick and click Adopt before they enter routine collection.

Exploration only proposes candidates and never starts collection by itself; you decide what joins the routine.

12. Protocol specs

Implementation details for SSH, Telnet, RDP, SFTP and the local shell — authentication matrix, option negotiation, channel capabilities, trust model and system requirements — are collected on the technical specs page and kept in sync with the desktop app.

13. Custom models (advanced, optional)

Not needed by default: cloud models work after sign-in. Configure this only when you want to use your own model service.

1. Open Home → Settings → Models → Add model.
2. Pick a vendor template (OpenAI, Anthropic, DeepSeek, Ollama, OpenRouter, Moonshot, Qwen, Zhipu) or a custom base URL.
3. Enter the API key and model name, save, and it becomes selectable in the Deck input bar.

Custom model keys are stored only on this machine; their usage goes to your own service and does not count against the subscription quota.

14. Safety and defaults

• The assistant defaults to Ask for approval. Get used to the tool-call rhythm with read-only tasks before widening the boundary.
• "Remember allow" only applies to similar commands; it is not a global mute.
• "Allow all" still blocks commands on the danger word list, which you can customise.
• Host fingerprints must be confirmed on first connection; when one changes, verify the change is expected before choosing how to trust it.
• Device passwords, key material and custom model keys are within the local trust boundary — protect the computer that runs Termini.
• Automatic alert repair (L2) follows your confirmation policy just like any assistant write action.

See the features page for a fuller description; after installing, follow steps 2 → 3 → 4 → 6 above to complete your first loop.