MCP Debugger

What the mcp-debugger MCP server does

The mcp-debugger MCP server gives an AI agent control of live debugging sessions through MCP tools. It can launch programs, attach to existing processes, pause execution, continue, step through code, inspect stack frames and variables, and evaluate expressions. Sessions can also expose program stdout and stderr through buffered output retrieval and a subscribable MCP resource.

Supported debugging environments cover Python, JavaScript and TypeScript, Ruby, Rust, Go, Java, .NET, C and C++, COBOL, and Dart/Flutter. The exact launch and attach options depend on the language adapter. For example, Python uses debugpy, Ruby uses rdbg, Go uses Delve, Java uses a JDI bridge, and C/C++ uses CodeLLDB. The Dart and Flutter workflows use the SDK-provided debug adapters rather than a vendored debugger.

How it works

The server translates MCP requests into Debug Adapter Protocol operations or language-specific attach workflows. It can run as a standalone Node.js process, so an IDE is not required. Remote attachment is available for several runtimes through a host and port, VM-service URI, JDWP, or similar connection; native C/C++ and COBOL targets can also be attached by process ID. Kubernetes workflows include port-forward attachment for interpreted runtimes and kubectl debug --target for native processes.

Breakpoints may be addressed by source location, matching statement content, or function name. An expectedContent check can validate that a line still contains the intended code, while content and function anchors can be resolved again after a restart. Logpoints use a logMessage template to record interpolated values without stopping execution. Sessions pause on uncaught exceptions by default, exposing the exception information, stack, and local state.

Setup and configuration

The package can be distributed with npm or npx, and a Docker image is available. It runs anywhere Node.js runs, including CI runners, containers, Kubernetes pods, SSH machines, and agent sandboxes. The server supports stdio and Streamable HTTP transports. Its HTTP transport is locked to loopback and checks Host and Origin headers.

Language-specific prerequisites vary. Java requires JDK 21 or newer, .NET uses netcoredbg, and Dart or Flutter debugging requires the corresponding SDK. JavaScript, C/C++, and COBOL debug engines ship with the package. Rust support uses CodeLLDB, with Windows requiring the GNU toolchain. The list_supported_languages tool reports availability and reasons for unavailable modes. A doctor command is also provided for environment checks.

Tools and capabilities

The mcp-debugger MCP server includes tools and capabilities for:

• Creating and controlling launch or attach sessions.
• Setting, listing, removing, and clearing breakpoints while a session runs.
• Restarting a session and reapplying its breakpoint configuration.
• Stepping, continuing, pausing, and inspecting stack frames and variables.
• Evaluating expressions in the current debugging context.
• Reading buffered debuggee output with a cursor.
• Listing supported languages and their available modes.
• Exposing a token-gated, read-only loopback DAP endpoint through expose_session for an IDE mirror.