Guides

Connect an AI agent to a shell with the RevSheller MCP server

Free, and no account needed for one session. Create a listener and follow along.

Create a listener

Authorized use only. Point an agent only at CTFs, your own labs and systems you have explicit permission to test. You are responsible for what your agent does.

What this gives you

RevSheller runs a remote Model Context Protocol server. An MCP client such as Claude Code or Claude Desktop can use it to create a session, get the command to run on your target, then send commands and read the output. You write no code and run no listener of your own.

Access: MCP is free for every registered account during the beta, and is planned as a Pro Plus feature afterwards. Plans are not final; see pricing.

Step 1: get an API key

Register, sign in, and generate a key on the account page. Keys start with rsh_ and are shown once. A new key replaces the old one. Treat it like a password and keep it out of shared configs and repositories.

Step 2: add the server

The endpoint is https://revsheller.com/mcp (Streamable HTTP, stateless, JSON responses). It authenticates with a Bearer API key. Cookies are not accepted.

Claude Code

claude mcp add --transport http revsheller https://revsheller.com/mcp \
  --header "Authorization: Bearer rsh_your_api_key"

Claude Desktop reaches remote servers through the mcp-remote bridge. Add this to claude_desktop_config.json:

{
  "mcpServers": {
    "revsheller": {
      "command": "npx",
      "args": ["-y", "mcp-remote@0.16.7", "https://revsheller.com/mcp",
               "--header", "Authorization: Bearer ${RSH_KEY}"],
      "env": { "RSH_KEY": "rsh_your_api_key" }
    }
  }
}

Restart the client and confirm the RevSheller tools appear.

Step 3: the workflow

  1. The agent calls create_session and receives a session id plus ready-to-run target commands for POSIX and Windows.
  2. You run that command on a machine you are authorized to test.
  3. The agent calls session_status until the target is connected.
  4. The agent calls run_command for each step and close_session when done.

The full set of tools is create_session, get_target_command, session_status, send_input, read_output, run_command, list_sessions and close_session. Input and output details are in the API docs.

A sensible first run

Start with a throwaway VM or a CTF box you already own, not anything you care about. Ask the agent something narrow, such as "create a RevSheller session, give me the Linux command, and once I have run it, report the kernel version and current user". That exercises the whole loop: create_session, your paste, session_status, then run_command. Once that works, you can hand it a longer task, keeping in mind that the agent acts with whatever access the shell has. Give it a scope in plain words, and review the commands it runs, because it is your authorization and your responsibility.

You can watch, and you stay in control

The agent never receives the operator secret. The browser can attach to the same session while the agent works, and both of you can type. If the agent heads somewhere you do not want, close the session from the web interface or tell it to stop.

Limits to know about

FAQ

Is MCP access free?
Yes for registered accounts during the beta; planned as Pro Plus afterwards.

Does the agent get my operator secret?
No. It is never returned over MCP.

Which clients work?
Any client supporting Streamable HTTP with a custom Authorization header. Claude Code connects directly; Claude Desktop uses mcp-remote.

Who is responsible for what the agent does?
You are. See the Terms of Service, and report abuse if you see misuse.

Ready to try it? One click creates a listener; you paste one command on a machine you are authorized to test.

Create a listener