mcp-jumpserver-gui-sucks

CAUTION: Operate production machines with extreme care. This MCP assumes no responsibility for production incidents caused by unsafe or incompetent model behavior.

A JumpServer 443-only MCP bridge for coding agents such as Codex and Claude. The project exposes a CLI-first, MFA-compatible, audit-preserving path into JumpServer assets without depending on port 2222 or any GUI-driven workflow in normal use.

Current Status

The main CLI and MCP chain is working against a real JumpServer instance:

• CLI-first login with terminal-entered MFA

• persisted durable access_key auth for REST discovery

• persisted authenticated web-session cookies for KoKo terminal flows

• asset, node, connect-method, and asset-access discovery

• KoKo 443 WebSocket probing

• one-shot remote command execution through KoKo

• managed multi-turn terminal sessions for MCP-driven shell interaction

• managed shell reuse for repeated command execution against the same asset/account target

• non-blocking buffered terminal output reads for managed sessions

• explicit managed-session command interruption with Ctrl-C

• verified effective-user tracking and managed sudo/su login-shell transitions

• process-local terminal idle reaping and session-cap enforcement

• explicit cookie-session refresh probing before terminal work

• a line-oriented CLI shell for non-MCP interactive terminal use

• managed KoKo elFinder SFTP sessions over port 443

• remote directory listing, metadata, directory creation, upload, download, rename, and deletion

• explicit overwrite and recursive-delete safety gates for file operations

The current implementation is usable, but it is not feature-complete yet. The most important known limitation is:

• terminal access still depends on a valid cookie-backed web session, so a fully expired terminal session still requires a fresh login run with MFA

Terminal-oriented entry points now accept either the concrete JumpServer account ID/alias required by the API or a user-facing account reference such as root, test-root, or the account username. The MCP resolves that reference to the concrete per-asset account ID before opening terminal sessions or creating connection tokens.

Related MCP server: Codex SSH Terminal MCP

Tracked Project Docs

• docs/live-instance-recon.md

• docs/web-terminal-flow.md

• docs/auth-state-format.md

• docs/cli-login-flow.md

• docs/sftp-flow.md

Upstream Reference Repositories

The repository keeps several untracked upstream JumpServer codebases under extern/ for protocol and behavior reference only. They are not runtime dependencies of this package.

• extern/jumpserver: backend API, authentication, and permission-model reference

• extern/koko: KoKo terminal gateway and WebSocket behavior reference

• extern/luna: legacy web-terminal frontend flow reference, especially around browser-driven terminal bootstrap behavior

• extern/lina: newer web UI and API usage-pattern reference

• extern/client: official client-side implementation reference for adjacent access workflows

Authentication Model

The runtime intentionally uses two auth layers:

• access_key for durable REST access

• authenticated web-session cookies for KoKo terminal access

Do not put live session secrets, cookies, or MFA values into MCP client config files. The intended flow is:

1. Run the CLI login command once.

2. Complete MFA in the terminal.

3. Let the tool persist auth state into the user-scoped application state directory.

4. Start the MCP server from Codex or Claude.

When the live JumpServer deployment enables a login captcha challenge, the CLI login command saves the captcha image under /private/tmp/ and opens it with the system image viewer before prompting for the captcha value in the terminal.

By default, persisted auth state lives under the platform-specific user application state directory:

• macOS example: ~/Library/Application Support/mcp-jumpserver-gui-sucks/auth-state.json

Advanced users can override the location with:

• MCP_JUMPSERVER_GUI_SUCKS_STATE_DIR

• MCP_JUMPSERVER_GUI_SUCKS_STATE_FILE

Install

Use the published package directly:

uvx mcp-jumpserver-gui-sucks --help

Login Before Starting MCP

uvx mcp-jumpserver-gui-sucks login \
  --base-url https://jumpserver.example.com \
  --username alice

Useful verification commands:

uvx mcp-jumpserver-gui-sucks doctor
uvx mcp-jumpserver-gui-sucks refresh-session --force

The login command persists state outside the repository. MCP client config should only describe how to find that state, not embed the secrets themselves.

MCP Configuration

The MCP server entrypoint is:

uvx mcp-jumpserver-gui-sucks serve

serve defaults to stdio, which is the correct transport for Codex and Claude desktop-style MCP clients.

Recommended Agent Terminal Workflow

When a coding agent plans to work on one machine for more than one command, the recommended workflow is:

1. Call jms_terminal_usage_guide.

2. Call jms_acquire_terminal_session with asset_ref and account_ref.

3. Call jms_get_terminal_identity before work where the effective operating-system user matters.

4. Call jms_switch_terminal_user to enter another user's login shell through sudo or su.

5. Use jms_run_terminal_command for short command-style work. Each result includes the verified current identity.

6. Use jms_send_terminal_input plus jms_read_terminal_output for shell-style interaction.

7. Call jms_interrupt_terminal_session when a command needs to be stopped.

8. Call jms_exit_terminal_user to leave one MCP-managed user shell and restore the previous user.

9. Call jms_close_terminal_session when the task is complete.

This keeps one KoKo shell open per target and avoids leaving many short-lived web-shell records behind in JumpServer.

jms_switch_terminal_user defaults to passwordless sudo. Set method="su" for target-account authentication, or provide the optional password argument when either method requires one. The MCP sends the password only after the remote terminal presents an authentication prompt, and it excludes the password from command text, results, and application logs. MCP clients may retain tool arguments in their own conversation or trace history, so treat a supplied password as exposed to the configured MCP client and model.

The managed session records the effective UID, username, home directory, shell, and user-switch depth. Prefer the explicit switch and exit tools over manually typing sudo su; ordinary command execution still refreshes the effective identity so manually initiated transitions remain visible.

Recommended Agent SFTP Workflow

1. Call jms_sftp_usage_guide.

2. Call jms_acquire_sftp_session with asset_ref and account_ref.

3. Reuse the returned session_handle for all file operations on that target.

4. Call jms_sftp_list or jms_sftp_stat before modifying remote paths.

5. Use jms_sftp_mkdir, jms_sftp_upload, jms_sftp_download, or jms_sftp_rename.

6. Call jms_close_sftp_session when the task is complete.

SFTP paths are absolute within the virtual root exposed by JumpServer, not necessarily the server's operating-system root. For example, if JumpServer maps an account's SFTP root to /tmp, the MCP path /project/file.txt refers to the server path /tmp/project/file.txt.

Uploads and downloads refuse destinations that already exist unless overwrite=true. KoKo's elFinder upload API does not expose an atomic create-if-absent operation, so an external client can still create the same remote path between the preflight check and upload. Avoid concurrent writers to the same path when overwrite=false is relied on as a safety gate. Remote deletion requires confirm=true. Because KoKo directory deletion is recursive, directories additionally require recursive=true; it is never enabled by default.

Codex (~/.codex/config.toml)

This matches the mcp_servers.* structure already used in your local ~/.codex/config.toml:

[mcp_servers.mcp-jumpserver-gui-sucks]
command = "uvx"
args = ["mcp-jumpserver-gui-sucks", "serve"]
startup_timeout_sec = 60.0
tool_timeout_sec = 600.0

[mcp_servers.mcp-jumpserver-gui-sucks.env]
MCP_JUMPSERVER_GUI_SUCKS_BASE_URL = "https://jumpserver.example.com"
MCP_JUMPSERVER_GUI_SUCKS_VERIFY_TLS = "true"
MCP_JUMPSERVER_GUI_SUCKS_USE_ENV_PROXY = "false"
MCP_JUMPSERVER_GUI_SUCKS_TERMINAL_IDLE_TIMEOUT_SECONDS = "3600"
MCP_JUMPSERVER_GUI_SUCKS_TERMINAL_REAP_INTERVAL_SECONDS = "30"
MCP_JUMPSERVER_GUI_SUCKS_MAX_TERMINAL_SESSIONS = "8"
MCP_JUMPSERVER_GUI_SUCKS_SFTP_IDLE_TIMEOUT_SECONDS = "3600"
MCP_JUMPSERVER_GUI_SUCKS_SFTP_REAP_INTERVAL_SECONDS = "30"
MCP_JUMPSERVER_GUI_SUCKS_MAX_SFTP_SESSIONS = "4"

# Optional when the default state directory is not desired.
# MCP_JUMPSERVER_GUI_SUCKS_STATE_DIR = "/Users/alice/Library/Application Support/mcp-jumpserver-gui-sucks"
# MCP_JUMPSERVER_GUI_SUCKS_STATE_FILE = "/Users/alice/Library/Application Support/mcp-jumpserver-gui-sucks/auth-state.json"
# MCP_JUMPSERVER_GUI_SUCKS_ORG_ID = "00000000-0000-0000-0000-000000000002"

tool_timeout_sec is a Codex-side MCP client setting. If it is omitted, Codex falls back to its own default per-tool timeout. Increase it when the agent may need to keep a single jms_* call open for longer-running terminal work. For terminal commands, pair it with a larger total_timeout_seconds on the specific jms_run_terminal_command or jms_execute_in_terminal_session call when needed.

Claude (~/.claude.json)

This matches the mcpServers JSON shape already present in your local ~/.claude.json:

{
  "mcpServers": {
    "mcp-jumpserver-gui-sucks": {
      "command": "uvx",
      "args": ["mcp-jumpserver-gui-sucks", "serve"],
      "env": {
        "MCP_JUMPSERVER_GUI_SUCKS_BASE_URL": "https://jumpserver.example.com",
        "MCP_JUMPSERVER_GUI_SUCKS_VERIFY_TLS": "true",
        "MCP_JUMPSERVER_GUI_SUCKS_USE_ENV_PROXY": "false",
        "MCP_JUMPSERVER_GUI_SUCKS_TERMINAL_IDLE_TIMEOUT_SECONDS": "3600",
        "MCP_JUMPSERVER_GUI_SUCKS_TERMINAL_REAP_INTERVAL_SECONDS": "30",
        "MCP_JUMPSERVER_GUI_SUCKS_MAX_TERMINAL_SESSIONS": "8",
        "MCP_JUMPSERVER_GUI_SUCKS_SFTP_IDLE_TIMEOUT_SECONDS": "3600",
        "MCP_JUMPSERVER_GUI_SUCKS_SFTP_REAP_INTERVAL_SECONDS": "30",
        "MCP_JUMPSERVER_GUI_SUCKS_MAX_SFTP_SESSIONS": "4"
      }
    }
  }
}

Supported Environment Variables

The current runtime reads these environment variables:

• MCP_JUMPSERVER_GUI_SUCKS_BASE_URL

• MCP_JUMPSERVER_GUI_SUCKS_ORG_ID

• MCP_JUMPSERVER_GUI_SUCKS_STATE_DIR

• MCP_JUMPSERVER_GUI_SUCKS_STATE_FILE

• MCP_JUMPSERVER_GUI_SUCKS_VERIFY_TLS

• MCP_JUMPSERVER_GUI_SUCKS_USE_ENV_PROXY

• MCP_JUMPSERVER_GUI_SUCKS_LOG_LEVEL

• MCP_JUMPSERVER_GUI_SUCKS_REQUEST_TIMEOUT_SECONDS

• MCP_JUMPSERVER_GUI_SUCKS_TERMINAL_IDLE_TIMEOUT_SECONDS

• MCP_JUMPSERVER_GUI_SUCKS_TERMINAL_REAP_INTERVAL_SECONDS

• MCP_JUMPSERVER_GUI_SUCKS_MAX_TERMINAL_SESSIONS

• MCP_JUMPSERVER_GUI_SUCKS_SFTP_IDLE_TIMEOUT_SECONDS

• MCP_JUMPSERVER_GUI_SUCKS_SFTP_REAP_INTERVAL_SECONDS

• MCP_JUMPSERVER_GUI_SUCKS_MAX_SFTP_SESSIONS

The recommended minimum MCP config is usually:

• MCP_JUMPSERVER_GUI_SUCKS_BASE_URL

• optionally MCP_JUMPSERVER_GUI_SUCKS_STATE_DIR or MCP_JUMPSERVER_GUI_SUCKS_STATE_FILE

Environment and operating-system proxy discovery is disabled by default so private JumpServer hosts are contacted directly. Set MCP_JUMPSERVER_GUI_SUCKS_USE_ENV_PROXY=true only when the JumpServer deployment is intentionally reached through the configured system proxy.

PyPI Release Automation

The repository now includes publish-pypi.yml.

Its behavior is intentionally:

• every push to main inspects pyproject.toml

• if the package version changed and that version does not already exist on PyPI, GitHub Actions builds and publishes it

• if the version did not change, the workflow skips publishing

• if the version already exists on PyPI, the workflow skips publishing

• workflow_dispatch can be used to publish the current version manually when it is not yet on PyPI

• after a successful PyPI publish, the workflow creates a matching v<version> GitHub Release with generated notes

• if PyPI already contains the current version but its GitHub Release is missing, the next workflow run backfills that Release without republishing the package

The publish job uses PyPI Trusted Publishing through GitHub OIDC. Configure PyPI to trust this repository and workflow before expecting the publish step to succeed.

Recommended PyPI trusted publisher settings:

• owner: ArtiPyHeart

• repository: mcp-jumpserver-gui-sucks

• workflow file: .github/workflows/publish-pypi.yml

• environment name: pypi

After Trusted Publishing is configured once, later pushes to main that bump project.version in pyproject.toml will publish to PyPI and create the matching GitHub Release automatically.

Current CLI Surface

• mcp-jumpserver-gui-sucks login

• mcp-jumpserver-gui-sucks paths

• mcp-jumpserver-gui-sucks doctor

• mcp-jumpserver-gui-sucks refresh-session

• mcp-jumpserver-gui-sucks resolve-target

• mcp-jumpserver-gui-sucks koko-probe

• mcp-jumpserver-gui-sucks terminal-exec

• mcp-jumpserver-gui-sucks terminal-shell

• mcp-jumpserver-gui-sucks save-state

• mcp-jumpserver-gui-sucks clear-state

• mcp-jumpserver-gui-sucks serve

Current MCP Tools

• jms_paths

• jms_status

• jms_terminal_usage_guide

• jms_sftp_usage_guide

• jms_profile

• jms_list_nodes

• jms_list_assets

• jms_get_asset

• jms_list_connect_methods

• jms_get_asset_access

• jms_resolve_terminal_target

• jms_list_connection_tokens

• jms_create_connection_token

• jms_expire_connection_token

• jms_refresh_terminal_auth

• jms_probe_koko_terminal

• jms_acquire_terminal_session

• jms_list_terminal_sessions

• jms_send_terminal_input

• jms_read_terminal_output

• jms_run_terminal_command

• jms_get_terminal_identity

• jms_switch_terminal_user

• jms_exit_terminal_user

• jms_interrupt_terminal_session

• jms_resize_terminal_session

• jms_close_terminal_session

• jms_acquire_sftp_session

• jms_list_sftp_sessions

• jms_sftp_stat

• jms_sftp_list

• jms_sftp_mkdir

• jms_sftp_upload

• jms_sftp_download

• jms_sftp_rename

• jms_sftp_delete

• jms_close_sftp_session

Operational Notes

• Managed terminal sessions are process-local and intended to live only for the MCP server process lifetime.

• Managed SFTP sessions are process-local, reused by asset/account pair, idle-reaped, and closed when the MCP process exits.

• Active SFTP operations are excluded from idle reaping, including transfers longer than the configured idle timeout.

• Terminal output is bounded in memory. Large streams retain their beginning and end and report truncation metadata.

• jms_sftp_upload accepts regular files from any local path; there is no local source-directory allowlist.

• Never set SFTP deletion flags speculatively. confirm=true requires approval of the exact path, and recursive=true requires explicit approval of recursive directory deletion.

• jms_terminal_usage_guide returns the preferred terminal workflow for coding agents and should be consulted at the start of terminal-heavy work.

• jms_acquire_terminal_session is the preferred high-level entrypoint for repeated work on one machine because it resolves the target and reuses an existing shell when possible.

• jms_run_terminal_command is the preferred path for short command execution on an already acquired session_handle.

• jms_get_terminal_identity verifies the effective UID and username before privileged work.

• jms_switch_terminal_user supports passwordless and password-backed sudo or su transitions while preserving the same JumpServer session.

• jms_exit_terminal_user exits one MCP-managed user shell and verifies that the previous identity was restored.

• jms_send_terminal_input plus jms_read_terminal_output are the preferred path for shell-style interaction and incremental polling.

• jms_interrupt_terminal_session is the supported way to stop a running managed-session command without throwing away the shell immediately.

• jms_interrupt_terminal_session accepts ctrl_c and the common alias SIGINT; both normalize to the same Ctrl+C behavior.

• The default managed shell idle timeout is 1 hour. Override it with MCP_JUMPSERVER_GUI_SUCKS_TERMINAL_IDLE_TIMEOUT_SECONDS if a different retention window is required.

• When the MCP server process exits normally, it closes all managed KoKo shells before returning.

• terminal-shell is line-oriented, not a full raw TTY emulator.

• Terminal entrypoints preflight the cookie-backed web session before opening KoKo.

• If the cookie-backed session is already invalid, terminal calls fail early with an explicit re-login requirement instead of a low-level websocket failure.

• REST discovery can continue to work when the durable access_key remains valid, even if terminal access requires a fresh login.