SSH Session MCP

ssh-session-mcp

中文 | English

[License: Apache%202.0] [Node.js Version] [TypeScript] [npm version]

Persistent shared-terminal runtime for MCP clients over SSH.

ssh-session-mcp gives the user and the AI the same SSH PTY session, adds a browser viewer, tracks who typed what, and makes long-running remote work manageable instead of stateless.

[ssh-session-mcp hero demo]

Contents

• Install At A Glance
• Project Structure
• Quick Start
• Docker Status
• MCP Tools
• Configuration Summary
• Security
• Docs
• Development

Install At A Glance

• Normal users do not need to git clone this repository.
• Preferred install path for MCP clients: npx -y ssh-session-mcp --viewerPort=auto
• Preferred install path for human operators who want local binaries: npm install -g ssh-session-mcp
• Official container distribution can be published to a public registry such as docker.io/zwawa/ssh-session-mcp
• git clone is only for contributors, source builds, and local development.
• For the common desktop MCP workflow, npx or a global npm install is still the lowest-friction path. Docker is mainly useful when you want a pinned runtime, container-based deployment, or registry-backed distribution.

Why It Exists

Most SSH-oriented MCP servers can execute commands, but they do not manage terminal state well enough for real collaboration.

ssh-session-mcp focuses on the missing runtime layer:

• One shared PTY for both the human and the AI
• Browser terminal for live inspection and manual intervention
• Input lock so the AI does not type over the user
• Safe/full execution modes for risky commands
• Configurable default policy rules plus session-level custom rule overrides
• Async command tracking for long-running remote work
• Multi-device and multi-connection profile support
• Local debug mode for demos, offline testing, and prompt iteration

Best Fit

• AI-assisted remote development on Linux boards and SSH servers
• Embedded, ROS, training, and deployment hosts that need a real terminal
• Users who want the AI to help, but do not want to surrender the terminal
• MCP Marketplace listings where the install and demo path must be clear

Project Structure

Key directories and files:

Quick Start

1. Agent-First Install (Auto-download on first run)

If the goal is to let Claude Code, Codex, or OpenCode install the server automatically, prefer npx -y ssh-session-mcp in the MCP command instead of a prior global install.

For Cline Marketplace and other agent installers, see llms-install.md. This repo is structured to be one-click installable through an npx -y ssh-session-mcp --viewerPort=auto command.

Claude Code

claude mcp add --transport stdio ssh-session-mcp -- npx -y ssh-session-mcp --viewerPort=auto

Windows note from the Claude Code docs: native Windows users should wrap npx with cmd /c for stdio MCP servers.

claude mcp add --transport stdio ssh-session-mcp -- cmd /c npx -y ssh-session-mcp --viewerPort=auto

Codex

codex mcp add ssh-session-mcp -- npx -y ssh-session-mcp --viewerPort=auto

OpenCode

OpenCode's opencode mcp add flow is interactive. Choose a local MCP server and use this command:

npx -y ssh-session-mcp --viewerPort=auto

If you prefer config instead of the interactive flow:

{  "$schema": "https://opencode.ai/config.json",  "mcp": {    "ssh-session-mcp": {      "type": "local",      "command": ["npx", "-y", "ssh-session-mcp", "--viewerPort=auto"]    }  }}

This is the closest thing to "automatic installation" for stdio MCP servers today: the MCP client stores the command, and npx -y downloads the package automatically the first time it runs.

2. Fastest Local Demo

npm install -g ssh-session-mcpssh-session-mcp-ctl launch --local --viewerPort=auto

This starts a local shell instead of SSH and opens the browser terminal, which is the easiest way to test the MCP runtime before touching a real server.

3. Register As An MCP Server

Use the MCP server binary directly when wiring a client:

# Global installnpm install -g ssh-session-mcp
# Server command used by MCP clientsssh-session-mcp --viewerPort=auto

# Claude Codeclaude mcp add --transport stdio ssh-session-mcp -- ssh-session-mcp --viewerPort=auto
# Codex CLIcodex mcp add ssh-session-mcp -- ssh-session-mcp --viewerPort=auto

If you prefer npx instead of a global install:

npx -y ssh-session-mcp --viewerPort=auto

4. Connect To A Real SSH Target

Create .env from .env.example:

cp .env.example .env

SSH_HOST=YOUR_DEVICE_HOSTSSH_PORT=22SSH_USER=YOUR_DEVICE_USERSSH_PASSWORD=SSH_KEY=VIEWER_PORT=autoAUTO_OPEN_TERMINAL=falseSSH_MCP_MODE=safe

Then launch:

ssh-session-mcp-ctl launch --viewerPort=auto

5. Multi-Device Config

For multiple boards or named targets, create ssh-session-mcp.config.json:

{  "defaultDevice": "DEVICE_A_ID",  "devices": [    {      "id": "DEVICE_A_ID",      "host": "DEVICE_A_HOST",      "port": 22,      "user": "DEVICE_A_USER",      "auth": { "passwordEnv": "DEVICE_A_PASSWORD" },      "defaults": {        "term": "xterm-256color",        "cols": 120,        "rows": 40,        "autoOpenViewer": true,        "viewerMode": "browser"      }    }  ]}

Discovery order:

1. --config=/path/to/config.json
2. Workspace ssh-session-mcp.config.json
3. User-global config
4. Legacy .env fallback

Important:

• Config discovery is based on the MCP process working directory.
• auth.password is intentionally unsupported. Use auth.passwordEnv or auth.keyPath.
• Secrets belong in .env or the parent environment, not in repo-tracked JSON.

6. Docker Status

Public Docker images should be distributed through Docker Hub, with GitHub Container Registry as an optional secondary registry:

docker.io/zwawa/ssh-session-mcp:<version>docker.io/zwawa/ssh-session-mcp:latestghcr.io/zw-awa/ssh-session-mcp:<version>

Recommended container launch for a real SSH target:

docker run --rm -i \  -p 8793:8793 \  -e VIEWER_PORT=8793 \  -e VIEWER_HOST=0.0.0.0 \  -e SSH_HOST=YOUR_DEVICE_HOST \  -e SSH_PORT=22 \  -e SSH_USER=YOUR_DEVICE_USER \  -e SSH_PASSWORD \  docker.io/zwawa/ssh-session-mcp:latest

Export the password in your shell first instead of placing it directly on the command line.

Recommended launch for profile-based config:

docker run --rm -i \  -p 8793:8793 \  -e VIEWER_PORT=8793 \  -e VIEWER_HOST=0.0.0.0 \  -e SSH_MCP_CONFIG=/workspace/ssh-session-mcp.config.json \  -v "$PWD/ssh-session-mcp.config.json:/workspace/ssh-session-mcp.config.json:ro" \  -v "/path/to/host/keys:/workspace/keys:ro" \  docker.io/zwawa/ssh-session-mcp:latest

Equivalent Compose example:

docker compose up -d

See docker-compose.yml for a ready-to-run example that mounts ssh-session-mcp.config.json, publishes the viewer on 8793, and uses SSH_KEY_DIR when set or falls back to a dedicated ./keys directory. For the full Docker guide, including the legacy .env compose variant and MCP client config snippets, see docs/docker.md. For a container-oriented profile example, see docs/examples/ssh-session-mcp.config.docker.example.json.

Container-specific notes:

• The image defaults VIEWER_PORT to 8793 when unset so the browser viewer can be published reliably.
• The image defaults VIEWER_HOST to 0.0.0.0 inside the container so the mapped port is reachable from the host.
• AUTO_OPEN_TERMINAL defaults to false in the container because browser auto-open from inside a container is usually not useful.
• Mount config files or SSH keys read-only when possible.
• Prefer mounting SSH keys from a directory outside the repo root.
• In docker-compose.yml, SSH_KEY_DIR overrides the default key mount path. If it is unset, Compose falls back to ./keys, not the repo root.
• Avoid putting passwords directly on the command line. Prefer exported env vars, Compose .env, or --env-file.
• For stdio MCP clients, Docker is viable, but host-native npx is still simpler unless your client explicitly prefers containerized commands.

Docker-based MCP client command examples:

# Claude Codeclaude mcp add --transport stdio ssh-session-mcp -- docker run --rm -i -p 8793:8793 -e VIEWER_PORT=8793 -e VIEWER_HOST=0.0.0.0 docker.io/zwawa/ssh-session-mcp:latest
# Codex CLIcodex mcp add ssh-session-mcp -- docker run --rm -i -p 8793:8793 -e VIEWER_PORT=8793 -e VIEWER_HOST=0.0.0.0 docker.io/zwawa/ssh-session-mcp:latest

For JSON-based MCP clients, the same pattern works by using docker as the command and passing the remaining run ... docker.io/zwawa/ssh-session-mcp:latest tokens as args.

This is useful when:

• The primary workflow is a local stdio MCP server command, not a long-lived network service.
• You want a pinned Node/runtime environment without a local install.
• You need registry-based distribution for a team or managed host.
• You want container-level isolation for the MCP server process.

For many users, publishing to npm and recommending npx -y ssh-session-mcp --viewerPort=auto is still the lower-friction install path.

Viewer And Collaboration Model

The browser viewer is not decorative. It is part of the workflow:

• The user can see exactly what the AI did.
• The AI can pause when the user takes over.
• Password prompts, pagers, and editors become visible state instead of hidden failure modes.
• Session diagnostics and history turn terminal debugging into something inspectable.

Marketplace-Friendly Flow

For users:

install -> launch viewer -> connect once -> keep the session alive -> let the AI help

For agents:

ssh-quick-connect -> ssh-run -> inspect output -> ssh-command-status if needed -> ssh-run again

Use AGENT.md when you want the AI to install, inspect config, connect devices, and help the user end-to-end. Compatibility notes for older agent setups remain in AI_AGENT_GUIDE.md.

Core Differences From A Stateless MCP SSH Wrapper

• Shared PTY instead of one-off command execution
• Actor-aware transcript markers for user, system, and agent input
• Terminal-state checks before dangerous or nonsensical writes
• Auto cleanup for sessions and viewer processes
• Session-scoped browser viewer with diagnostics and history
• Local debug mode with --local for offline testing

Operation Modes

Each session now owns its own safe / full mode. Switching one browser terminal to full does not change other sessions.

The default rule set can be customized if needed. Custom rules now support:

• error: block the command
• warning: allow but surface a warning
• log: allow and annotate only

Rule precedence is error > warning > log, and within the same level, earlier rules win.

Lock Policy

The browser terminal UI lets the operator choose one of these input policies:

When the terminal is not available for agent input, tools such as ssh-run, ssh-session-send, and ssh-session-control return a blocked response instead of forcing input into the PTY.

MCP Tools

Recommended Daily Tools

Full Tool Catalog

Local Operator Commands

These helpers are for humans on the workstation that owns the viewer:

ssh-session-mcp-ctl statusssh-session-mcp-ctl devicesssh-session-mcp-ctl launch --viewerPort=autossh-session-mcp-ctl launch --local --viewerPort=autossh-session-mcp-ctl logs --tail=60ssh-session-mcp-ctl cleanup

Default rule library management for operators:

ssh-session-mcp-config policy list --scope=mergedssh-session-mcp-config policy set error-kubectl-delete --pattern="\\bkubectl\\s+delete\\b" --category=dangerous --action=error --priority=0 --message="kubectl delete is blocked in safe mode"ssh-session-mcp-config policy remove error-kubectl-delete

Equivalent repo-local commands also exist:

npm run launchnpm run statusnpm run devicesnpm run logsnpm run cleanup

Configuration Summary

Key environment variables:

Distributed v0

Distributed v0 intentionally implements a narrow boundary:

• Supported runtime modes: single-node and distributed
• Distributed mode shares control-plane state only: node heartbeat, session metadata, binding metadata, command metadata, and viewer access policy
• When the current replica is not the owner, HTTP APIs return REMOTE_OWNER, HTML pages render a remote-owner error page, and websocket attaches close with code 4009
• Cross-node PTY migration is not supported
• Transparent cross-node HTTP or websocket proxying is not supported

Distributed v0 requires Redis for real multi-node deployments. SSH_MCP_STORE=memory only exists for local skeleton testing and does not provide a shared store across replicas.

Distributed configuration:

Recommended distributed env example:

SSH_MCP_RUNTIME_MODE=distributedSSH_MCP_STORE=redisSSH_MCP_REDIS_URL=redis://redis:6379/0SSH_MCP_NODE_ID=node-aSSH_MCP_PUBLIC_BASE_URL=https://ssh-mcp.example.comSSH_MCP_AUTH_MODE=proxySSH_MCP_TRUST_PROXY=trueSSH_MCP_AUTH_USER_HEADER=x-forwarded-userSSH_MCP_AUTH_ROLE_HEADER=x-forwarded-role

Proxy auth is most useful in distributed mode behind a trusted reverse proxy. The built-in role mapping is:

• viewer_read: pages, session list/read endpoints, history, diagnostics, health, readiness, metrics
• viewer_write: attach input, resize, control
• session_admin: mode changes, policy updates, close, set-active, debug-agent actions, local debug session creation

Macro / Environment Variable Reference

Use these variables according to your installation path:

Minimum Required Settings

Choose one of these minimum configuration sets:

• Local demo: SSH_MCP_LOCAL=true and VIEWER_PORT=auto
• Legacy SSH with password: SSH_HOST, SSH_USER, SSH_PASSWORD
• Legacy SSH with key: SSH_HOST, SSH_USER, SSH_KEY
• Profile-based mode: ssh-session-mcp.config.json, plus any passwordEnv variables referenced by that config
• Docker Compose profile mode: ssh-session-mcp.config.json, optional SSH_KEY_DIR, optional SSH_SESSION_MCP_IMAGE

Container Runtime Notes

• Container defaults now set SSH_MCP_LOG_MODE=stderr so logs go to the container runtime without corrupting stdio MCP transport.
• Mount SSH_MCP_STATE_DIR persistently when you want viewer policy, server info, and state files to survive container restarts.
• Distributed multi-node deployments need Redis plus a routable SSH_MCP_PUBLIC_BASE_URL per replica.
• Distributed v0 does not provide cross-node PTY migration or transparent cross-node proxying. Route requests to the owner node when you receive REMOTE_OWNER.
• Health endpoints:
  • /livez for process liveness
  • /readyz for readiness checks
  • /metrics for Prometheus text metrics
• Example single-instance Kubernetes baseline: docs/examples/ssh-session-mcp.k8s.single-instance.yaml
• Example distributed Kubernetes baseline: docs/examples/ssh-session-mcp.k8s.distributed.example.yaml
• Primary Kubernetes installation path: deploy/helm/ssh-session-mcp

Example config file: docs/examples/ssh-session-mcp.config.example.json

Security

• The package never requires raw passwords inside tracked JSON config.
• .env is ignored by git and npm.
• Viewer HTTP binds to localhost by default.
• The MCP server treats terminal mode and input lock as first-class safety signals.
• CI runs Trivy filesystem and container-image scans against high and critical vulnerabilities.
• CI installs a pinned Trivy CLI release with checksum verification instead of relying on a floating third-party action tag.
• Release builds generate a CycloneDX SBOM for the published GHCR image digest and attach it to the GitHub release.
• Release builds sign the published GHCR image digest with keyless Cosign.
• GHCR digest is the primary verification path. Docker Hub remains a distribution path, not the main signature-verification target.

See SECURITY.md for the full policy.

Platform Notes

• Windows 10/11: first-class host environment
• Linux: strong fit for headless MCP + browser viewer workflows
• macOS: standard Node.js path supported
• Remote Linux hosts: first-class target

More detail: docs/platform-compatibility.md

Docs

• AGENT.md
• AI_AGENT_GUIDE.md
• llms-install.md
• docs/contracts.md
• docs/failure-taxonomy.md
• docs/acceptance-scenarios.md
• docs/docker.md
• docs/kubernetes.md
• docs/ingress-proxy-auth.md
• CHANGELOG.md

Development

Clone the repo only if you want to modify the source, run tests locally, or build release artifacts.

npm installnpm run buildnpm run testnpm run validate:reponpm run build:site

GitHub Actions included in this repo can:

• run CI on push and pull request
• deploy a GitHub Pages landing page from dist/
• build a tagged GitHub Release with the npm package tarball attached

License

Apache-2.0. See LICENSE.