Last updated July 2, 2026

MCP Tailnet Manager

MCP Tailnet Manager is a desktop app that opens a local MCP bridge for any of your servers, so an AI coding agent — Claude Code, Codex, Cursor, Windsurf, OpenCode — can work on the host directly over SSH. No cloud round-trip, no HostAtlas agent on the server, no third party sees your commands or files.

Built with Tauri 2 + React 19 + Rust, mirroring the SSH Client stack.

What it does

Keep a local server list — add hosts by IP or hostname, or discover them via Tailscale if it’s installed locally
Open an MCP server for a chosen host — the app tests the SSH connection with your local key, starts a local MCP bridge process, and registers the allowed tools / paths / commands for that host
Point your agent at the local endpoint — commands run on the remote over your direct SSH connection

Install

Download the installer for your platform from the GitHub Releases page or from install.hostatlas.app:

Platform	File
macOS (Apple Silicon)	`.dmg` bundle at `install.hostatlas.app/mcp-tailnet/latest/`
macOS (Intel)	`.dmg` bundle
Linux (x86_64)	`.deb` or `.AppImage`
Windows (x86_64)	`.msi` installer

Each release also ships a standalone mcp-tailnet bridge binary per platform — that’s the process AI agents launch to actually talk to a host (see below).

The top-level https://install.hostatlas.app/mcp-tailnet/version.json carries the full list under binaries[] (filename, size, sha256, url) and both folders publish a files.json sidecar for checksum verification.

Secure by design

The app never touches key material. It does not generate, store, deploy, or sync SSH keys, passphrases, or passwords. It uses the SSH keys already installed locally:

~/.ssh — private keys
ssh-agent — running key agent
~/.ssh/config — host-specific settings
~/.ssh/known_hosts — host-key verification

If the local ssh command works for a host, so does the app. With Tailscale, connect over the tailnet IP or hostname just like you would with any other SSH client.

The MCP bridge (`mcp-tailnet`)

mcp-tailnet is a standalone binary that speaks the MCP protocol (JSON-RPC 2.0 over stdio) and proxies every tool call to a remote host over direct SSH — using the system ssh binary with a shared ControlMaster connection for speed.

An AI agent launches it on demand:

mcp-tailnet serve --server <host>

Available tools

Tool	Description
`list_directory`	List a directory on the remote host
`read_file`	Read a file’s contents
`write_file`	Write a file (subject to `require_approval` gate)
`search_files`	Search recursively for a pattern in files
`execute_command`	Run a shell command (whitelist-enforced)
`git_status`	Get `git status` output from a repository on the host

Enforcement (in the bridge itself, not in the app)

Global forbidden denylist — rm -rf /, mkfs, shutdown, and other catastrophic commands are always blocked, no override
allowed_paths — file operations must stay within the configured directories
allowed_commands — execute_command is checked against the per-host whitelist
require_approval — blocks write_file calls until the user approves them, when set

Enforcement lives in the bridge process, not in the app UI. Killing the app doesn’t unlock the bridge, and the bridge is what your AI agent actually talks to.

Put the bridge on your `PATH`

Every release publishes mcp-tailnet-<os>-<arch> alongside the app bundle. Put it somewhere your AI agent can find it:

# macOS / Linux — from install.hostatlas.app
sudo curl -fsSL https://install.hostatlas.app/mcp-tailnet/latest/mcp-tailnet-$(uname -s | tr A-Z a-z)-$(uname -m | sed 's/x86_64/amd64/;s/aarch64/arm64/') \
  -o /usr/local/bin/mcp-tailnet && sudo chmod +x /usr/local/bin/mcp-tailnet

The Agents tab in the app gives you the exact config snippet per client — Claude Code, Codex, Cursor, Windsurf, or OpenCode — with the right path and args pre-filled.

Status

Standalone-only for v1. You add servers, connect, open the MCP bridge, and talk through it. HostAtlas login and server-list sync are deferred to a later release (which will move to Akyros ID SSO). The app UI shows a “Standalone” state until then.

The MCP bridge speaks the protocol and executes tools over SSH today. Audit-log capture from live tool calls is the next piece on the roadmap.

Requirements

macOS 11+, Windows 10+, or a modern Linux distribution with WebKitGTK 4.1
The system ssh command available on your PATH
(Optional) Tailscale locally installed if you want the tailnet-discovery feature

SSH Client — the human-facing sibling app; same stack, terminal-first
MCP Server — exposes the HostAtlas platform API (not per-host SSH access) to your AI agent

Was this page helpful?

Report issue ↗