Hybrid Access: AI on SSH servers without an agent
With Hybrid Access an SSH server you already use in WECOMMANDER can be used by AI without installing anything on it. The connection is made by WECOMMANDER on your own computer, so your SSH keys and passwords never leave it. Connection method and AI access are separate choices: no AI sees a server until you turn AI access on for it.
Where it is available
| Windows app | macOS app | Linux Management Center | |
|---|---|---|---|
| Remote AI (ChatGPT, Claude.ai through wecommander.app) | Yes | Yes | Yes |
| AI on this computer (local MCP) | Yes | No settings page yet | Not yet |
Turn it on for a server
- The server must use SSH (not Telnet, serial or Mosh).
- Connect once and approve its host key. AI never trusts a server on first use, and a changed key is refused.
- It must sign in without asking: a saved password, a key that opens without a typed passphrase, or this computer's own keys or SSH agent. The AI can never type a password.
- Edit the server → AI access, and turn on “Remote AI” (and on Windows, “AI on this computer” if you use local AI clients).
- Choose the permission preset and the sudo setting.
Presets and sudo
| Setting | Meaning |
|---|---|
| Read only | The AI can read files, list folders and look at the server's state. It cannot run commands or change anything, whatever its trust level. |
| With approval (recommended) | Reading is free; commands, writes, deletes and server changes need a single-use approval showing the exact command and server, unless the server's AI permissions allow it. |
| Timed autonomous (local only) | For local AI clients only: no approvals until the time ends (at most 8 hours), then back to With approval. Never applies to remote AI. |
| Sudo: never (default) | Commands containing sudo, su, doas, pkexec or run0 are refused. A sudo password is never stored or given to AI. |
| Sudo: sudoers | Only what the server's sudoers already allows without a password can run with sudo; you decide on the server. |
How a remote call works
- Your app publishes only a name, the preset and the sudo setting for each shared server; never the address, user or credentials. The AI sees a target id such as ssh-<24 hex characters>.
- The AI calls a tool with that target. The gateway checks the role, the read-only preset, the safety net, run as root, and whether an approval is needed.
- Your app picks the call up, checks everything again, runs it over SSH or SFTP, and sends back the result with secrets masked.
- If your computer or the app is off, the call fails at once with connector_offline (after 20 seconds without pickup, or 45 seconds without contact); nothing waits to run later.
Tools on SSH servers
process_exec, fs_list, fs_read, fs_stat, fs_write, fs_mkdir, fs_move, fs_delete, server_overview, server_disk_usage, server_processes, server_cleanup_scan, server_cleanup_run, server_process_signal, server_timers_list, docker_containers_list, docker_container_logs and docker_container_action. Other tools answer unsupported_for_target. Remote reads and writes are limited to 1 MiB each (4 MiB for local AI clients).
AI permissions of SSH servers
Each shared server has the same AI permissions as an agent device, on the web Devices page under SSH servers: trust level (strict, standard, full), run as user or root, the safety net and an allowlist of routine commands. Only the owner of the computer the server goes through can see and change them. Root needs both the computer's sudo setting (sudoers) and the owner's choice in the dashboard; then process_exec and fs_write run through sudo -n.
What it does not protect
- There is no folder limit over SSH: a command, and the file tools, reach everything the SSH user can. For a tighter limit install the agent, or use a separate SSH user with only the rights needed.
- The sudo guard is a guard, not a sandbox; the real limit is the SSH user's rights and the server's sudoers.
- Secret masking is best effort.
Always on
- The approval screen always shows the exact command.
- Passwords, keys and tokens in output are masked.
- Every call is in the audit log (tool, server, decision); never command text, file contents or output on the gateway.
- The AI can never change AI access, presets or sudo settings: no tool for that exists.
- AI permissions: trust levels, allowed commands, root and the safety net: Set per device and per SSH server what AI may do without asking: strict, standard or full trust (time-limited or not), allowed command patterns, run as root with two keys, the safety net and autonomous mode.
- AI access with MCP: how it works: What MCP (Model Context Protocol) is, the one WECOMMANDER MCP URL (https://wecommander.app/mcp), how a tool call reaches your machine, and the two ways to connect a server: the agent or SSH.
- Local MCP: without an account or through the desktop app: Run WECOMMANDER as a local MCP server over stdio for Claude Desktop, Cursor, VS Code or Codex: the agent's mcp-stdio mode with no account, or the Windows app's local endpoint for saved SSH servers.
- MCP tool reference: Every WECOMMANDER MCP tool, grouped: what it does, its risk and approval rule, where it works (agent devices, SSH servers, Linux only) and its key arguments.