sudo-proxy MCPサーバー

sudo-proxy

An MCP server that lets an agent run privileged, mutating commands — locally or over SSH — with a human keypress required on every one and no credential ever stored.

Privileged command execution proxy with an MCP server for AI agent integration. Receives requests over a Unix socket, shows a single-keypress TUI prompt for human approval, then escalates via sudo. Configure sudo-proxy-mcp in Claude Code or any MCP client and the model can run privileged commands — with explicit human approval on every one.

⚠️ Not yet independently audited. The security assurance in this repo is self-produced. Security reviewers: REVIEWING.md is a 15-minute on-ramp — the falsifiable claims, the trust boundary, and where we have no assurance. Reporting: SECURITY.md.

Architecture

AI model ──► sudo-proxy-mcp ──► Unix socket ──► sudo-proxy ──► TUI Y/N ──► sudo
             (MCP server)         │
                                  │
                            local socket, or SSH tunnel
                        (start_server spawns sudo-proxy --host
                         which sets up the tunnel and remote
                         server)

The TUI prompt asks for approval (single keypress), then sudo handles privilege escalation. The password prompt appears in the same terminal. This flow is identical for local and remote hosts. A non-privileged mode runs commands directly as the current user, still behind the same Y/N gate — see docs/usage.md.

Why not just use the Bash tool?

	Bash tool	sudo-proxy MCP
Privilege escalation	Not possible	sudo with human approval
Human review	None — executes immediately	TUI Y/N gate on every command (privileged and unprivileged)
Timeout	Up to 10 min, no user prompt	60 s TUI prompt + configurable overall timeout
Remote hosts	Not supported	SSH tunnel with TUI on remote terminal
Environment	Inherits shell env	Sanitized allowlist only
Audit trail	None	Server logs each request (with -v)

sudo-proxy fills the gap when a model needs to install packages, edit system files, manage services, or run any other command — with the human always in the loop, even when Claude Code is run with --dangerously-skip-permissions.

It is deliberately not a way to run unprivileged commands with less scrutiny than the Bash tool: an unprivileged command targeting the local machine is refused and delegated back to the Bash tool (which already applies your permission rules), and loopback aliases like 127.0.0.1 cannot dodge that. sudo-proxy handles privilege escalation and commands on remote hosts; there, unprivileged commands are gated the same way, and the per-command gate is only relaxed if you deliberately mark a host eligible in hosts.json and confirm once per session (a grant that is never persisted). See REVIEWING.md invariant G7.

For how this relates to mcp-firewall, sandboxing, polkit, doas, and other neighboring tools, see docs/comparison.md.

Features and non-features

What sudo-proxy does:

• Per-command human approval — a single-keypress Y/N TUI gate on every command, privileged and unprivileged, with no way to bypass it.
• Real privilege escalation via sudo — installs packages, edits system files, manages services; not limited to read-only diagnostics.
• Local and remote over one flow — the same approval TUI whether the command runs on this machine or on a remote host over an SSH tunnel.
• Stores no secret — the sudo password is typed live into the terminal; nothing is cached, encrypted-at-rest, or written to disk.
• Explicit argv — commands are passed and displayed exactly as they run, with no shell-string interpolation.
• Sanitized environment — a fixed allowlist, not the inherited shell environment.
• Audit trail — each request is logged by the server (-v).
• Works under --dangerously-skip-permissions — the human gate holds even when the agent's own permission prompts are disabled.

What sudo-proxy deliberately does not do:

• No stored or managed credentials — it is not a password cache or a secrets manager.
• No unattended execution — there is no auto-approve mode; a human approves each command or it does not run.
• Not a sandbox — it grants real privilege rather than isolating or faking it.
• Not a policy engine or ACL — the human at the keypress is the policy; there are no rules to write or maintain.
• Not read-only — it is not restricted to a whitelist of safe diagnostic commands.
• No hosted service — it runs on your own machine; commands never transit a third-party relay.
• Not a general remote shell — no persistent interactive sessions, SFTP browser, or fleet manager; just gated one-shot commands.

Quickstart

Install the binaries (needs a Rust toolchain; prebuilt static binaries are on Releases):

cargo install sudo-proxy

Or fetch the prebuilt static binaries without compiling, via cargo binstall:

cargo binstall sudo-proxy

cargo binstall downloads the release tarball from GitHub and verifies its minisign signature (public key RWT7gwtBU0v4puI76u0oYwMAT9nmYwGimSOnqJJ+kHExsjTDQj1eZkMW, key ID A6F84B53410B83FB) before installing; a missing or bad signature aborts the install. Prefer cargo install if you'd rather build from source.

Point your MCP client at the server — add to the project's .mcp.json or ~/.claude/claude_desktop_config.json:

{
  "mcpServers": {
    "sudo-proxy": {
      "command": "sudo-proxy-mcp"
    }
  }
}

Install as a Claude Code plugin

The repo is also a Claude Code plugin and its own marketplace. Instead of editing .mcp.json by hand:

/plugin marketplace add tarides/sudo-proxy
/plugin install sudo-proxy@tarides

The plugin registers the sudo-proxy MCP server and ships a skill that tells the model to reach for sudo-proxy only for privileged or remote work, to start the server before executing, and to write a clear description for each command so the approval prompt is easy to judge. It does not install the sudo-proxy-mcp binary — that must already be on your PATH (see the cargo install / cargo binstall steps above). A tool whose whole point is explicit human consent should not install itself silently.

Run commands

Then the model starts a server (opening a terminal with the approval TUI) and runs commands through it:

start_server()
execute({"argv": ["apt", "install", "nginx"], "description": "Install nginx"})

Each execute shows the exact command in the TUI; press y to approve, N (default) to deny. For remote hosts, pass host to start_server and execute.

When a prompt appears, the TUI rings the terminal bell and blinks the question line so a request is not missed while the daemon's terminal is unfocused or off-screen. The bell honors the terminal's own visual-bell/urgency setting (audible beep, taskbar/dock flash, or both). Both cues are on by default; set SUDO_PROXY_NO_BELL to silence the bell and SUDO_PROXY_NO_BLINK to stop the blink (same opt-out style as NO_COLOR).

MCP Registry

Listed in the official MCP Registry as mcp-name: io.github.tarides/sudo-proxy.

Documentation

• docs/install.md — install variants, remote deploy, building from source
• docs/mcp.md — MCP tools (start_server, execute, status, stop_server, update_host), config, known hosts
• docs/usage.md — CLI flags, non-privileged mode, SSH tunnels, agent forwarding
• docs/protocol.md — JSON-line wire protocol over the Unix socket
• docs/security.md — security model; docs/security-audit.md — point-in-time audit; docs/threat-model.md — STRIDE + attack tree; docs/formalisation-roadmap.md — graduated-assurance plan; docs/assurance-case.md — GSN argument
• proofs/ — machine-checked models: TLA+/PlusCal approval state machine (TLC, Rung 4) and ProVerif SSH-channel model (Rung 4); Kani bounded proofs live in src/proofs.rs (Rung 3)
• docs/pkexec.md — --pkexec mode and polkit auth caching (not recommended)
• docs/architecture.md — source layout and implementation status
• docs/comparison.md — how sudo-proxy relates to other tools

AI-assisted development

sudo-proxy was developed with substantial AI assistance using Claude Code. Commits where AI contributed materially carry a Co-Authored-By trailer naming the specific model. All design decisions, threat modeling, and the final form of every committed change were reviewed and approved by the human maintainer, who takes responsibility for the codebase.

This disclosure is provided in line with emerging industry practice around transparency about generative-AI involvement in software development. It is not a statement that the project is "AI-generated": the human-in-the-loop principle that the tool itself enforces at runtime is also the principle under which it was built.

License

MIT