swing-mcp

Let AI assistants operate Java Swing desktop applications — snapshot the UI, click, type, fill forms, read tables and trees, drive menus and dialogs. Inspired by chrome-devtools-mcp, but targeting any Swing UI instead of HTML pages.

A huge amount of business software is Java desktop software — internal tools, ERP clients, point-of-sale, lab and logistics systems — with no API and no web UI. Swing MCP gives that software a safe, permissioned door into the AI era: assistants like Claude can see the interface and act in it, without changing the application itself.

Quick start (Claude Desktop)

1. Make sure JDK 21+ is on your PATH.
2. Download swing-mcp.mcpb from the latest release.
3. Open the file — Claude Desktop installs it as an extension.
4. Ask Claude to launch_app your Swing application (or attach_to_app a running one by PID) and take it from there.

Using another MCP client? One guide and a one-minute video per client — Claude Code, VS Code + Copilot, Cursor, Devin Desktop (Windsurf), IntelliJ IDEA, Codex CLI, Gemini CLI: docs/installation.md · videos. Prefer the long version? 3-minute install-and-first-use video.

Need this connected to your own application — or an MCP connector for other software your business runs? CrossTech builds them: crosstech.solutions/swing-mcp.

Modules

• swing-mcp-server — Spring Boot MCP server (stdio transport) exposing Swing automation tools.
• swing-mcp-agent — Java agent loaded into the target Swing JVM (at launch via -javaagent, or dynamically by PID). Runs a localhost-only JSON line-protocol socket server that executes commands on the Swing EDT.
• swing-mcp-common — Shared command/DTO types between server and agent.
• swing-mcp-demo — Demo Swing application used for integration testing.

How it works

MCP client (stdio) ──▶ swing-mcp-server ──localhost socket──▶ swing-mcp-agent (inside target JVM) ──▶ Swing EDT

1. The MCP client calls launch_app (starts a JVM with the agent preloaded) or attach_to_app (loads the agent into a running JVM by PID).
2. The agent binds a loopback-only port in swing.mcp.agent-port-min..max and reports it back through a response file.
3. Tools such as take_snapshot, click, and fill are forwarded as JSON line commands and executed on the Event Dispatch Thread.

See docs/tools for the full tool documentation (per-category pages), or docs/tool-reference.md for the single-page quick reference.

Skills

The skills directory contains agent skills (SKILL.md files) that teach AI coding agents how to use Swing MCP effectively — core workflows, UI testing patterns, and troubleshooting.

Building

Requires JDK 21+ and Maven.

mvn verify

GUI integration tests are skipped in headless environments; CI runs them under xvfb.

Running

Build everything (or unzip the released swing-mcp.mcpb — the two jars are in its server/ folder), then register the server with your MCP client (see docs/installation.md for per-client instructions — Claude Desktop, Claude Code, VS Code, Cursor, Devin Desktop, IntelliJ IDEA, Codex CLI, Gemini CLI):

{
  "mcpServers": {
    "swing": {
      "command": "java",
      "args": ["-jar", "/path/to/swing-mcp-server.jar"],
      "env": {
        "SWING_MCP_AGENT_JAR": "/path/to/swing-mcp-agent.jar"
      }
    }
  }
}

Try it against the demo app:

1. launch_app with java -jar swing-mcp-demo/target/swing-mcp-demo-1.2.2.jar
2. take_snapshot to discover component UIDs
3. click, fill, select_option, … to interact

MCP Registry

This server is published to the MCP Registry as io.github.crosstech-solutions-bv/swing-mcp, distributed as an MCPB bundle attached to GitHub releases. See docs/registry-publishing.md for how publishing works.

Security notes

• The agent listens on the loopback interface only.
• evaluate_java (arbitrary code execution in the target JVM) is disabled by default; enable with swing.mcp.evaluate.enabled=true.

See SECURITY.md for the full security model and how to report vulnerabilities.

Contributing

Contributions are welcome — see CONTRIBUTING.md for the PR workflow. Release history lives in CHANGELOG.md.

About

Swing MCP is built and maintained by CrossTech, an AI-first software studio. We build MCP connectors that plug businesses' existing software — web, cloud, and legacy desktop — into AI assistants: crosstech.solutions/mcp.

License

MIT