.NET Debugger for MCP

DotnetDebugger.Mcp

An MCP server that lets an AI assistant debug .NET programs: launch or attach to a process, set breakpoints, step through code, inspect variables and evaluate expressions. It also carries a searchable reference on how .NET debuggers work internally.

Published on NuGet as DotnetDebugger.Mcp. Built on clrdbg, driven over the Debug Adapter Protocol. The debugger travels inside the package and runs as a child process, so a crash in native debugging code costs the debug session rather than the server.

Origin. A fork of decriptor/SharpDbg.MCP, now far enough from it to carry its own name: the debugger layer speaks the Debug Adapter Protocol instead of calling a debugger's internal API, the debugger underneath has since changed from SharpDbg to clrdbg, launching a program under the debugger is new, and the two have not shared a commit since. This repository was called SharpDbg.MCP until the first release, and GitHub redirects the old name. The SHARPDBG_* settings were renamed to DOTNET_DEBUGGER_MCP_* in the release that followed 0.6.3; the old names are no longer read.

What it can do

• Attach to a running .NET process, or launch one and stop before its first line executes
• Set breakpoints by file and line or by method name, with conditions and hit counts
• Step over, into and out of code, and read the call stack for any thread
• Inspect local variables, expand objects member by member, and evaluate C# expressions
• Break on exceptions - all of them, only the unhandled ones, or only the types you name - and read what was thrown
• Read the debuggee's stdout and stderr
• Debug more than one process at once, each with its own breakpoints and stops
• Debug a .NET MAUI app on an Android emulator or phone, an iOS simulator or device, or this Mac as a Mac Catalyst app - building it and stopping before its first line, same as a desktop program
• Search embedded documentation on ICorDebug, the Debug Adapter Protocol and expression evaluation

What it cannot do

Some of these need changes in the underlying debugger rather than here.

• Report the exit code of a process it attached to. exit_code is null there. The debugger reads the code off a process it started itself and has none for one it was pointed at, so for an attached process there is nothing to report — and the protocol cannot say "unknown", only 0, which would be a number this server made up. A program started with launch_program does report its real code.
• Debug a self-contained single-file publish. The runtime is packed inside the executable, so the debugger shim cannot find it to load the matching components. The attempt fails immediately with CORDBG_E_DEBUG_COMPONENT_MISSING (0x80131C3C), reported as Attempting to register for runtime startup failed: -2146231236. Self-contained on its own works, and single-file on its own works; only the combination does not.
• Watch expressions, hot reload and data breakpoints are not implemented.

Requirements

• .NET 10 SDK or later
• An MCP-compatible client, such as Claude Code or Claude Desktop
• Windows, macOS or Linux

Install

One package carries the debugger and its native shims for every platform, so the same configuration works everywhere. There is nothing to clone or build.

Claude Code, for the current project:

claude mcp add dotnet-debugger -- dotnet tool exec DotnetDebugger.Mcp --version 0.8.0 --yes

Add --scope user to make it available in every project, or --scope project to write a .mcp.json that is committed and shared with your team.

VS Code, in .vscode/mcp.json for the workspace, or with MCP: Open User Configuration for every workspace. VS Code calls the section servers, not mcpServers:

{  "servers": {    "dotnet-debugger": {      "type": "stdio",      "command": "dotnet",      "args": ["tool", "exec", "DotnetDebugger.Mcp", "--version", "0.8.0", "--yes"]    }  }}

Claude Desktop, by editing ~/.config/Claude/claude_desktop_config.json on macOS and Linux, or %APPDATA%\Claude\claude_desktop_config.json on Windows:

{  "mcpServers": {    "dotnet-debugger": {      "command": "dotnet",      "args": ["tool", "exec", "DotnetDebugger.Mcp", "--version", "0.8.0", "--yes"]    }  }}

dnx DotnetDebugger.Mcp --version 0.8.0 --yes is the shorter equivalent, and the form NuGet.org suggests. Prefer dotnet tool exec in a client launched from a desktop environment rather than from a shell: dnx lives in the SDK directory, which such a client often does not have on its PATH, while dotnet reliably is.

A client only connects its MCP servers when a session starts, so restart it after changing the configuration. To confirm the server is there, run claude mcp list or ask the client to list .NET processes.

Why the version is pinned

Every command here names a version. Without --version, dotnet tool exec resolves the package on each start, so a newly published release would run on the next session without anyone having chosen it. A debugger can read and change everything in the process it attaches to, so which build of it runs should be a decision rather than a side effect. Change the version in the command to upgrade; the latest is on nuget.org.

Installing the tool once instead avoids the resolution step on every start:

dotnet tool install -g DotnetDebugger.Mcp --version 0.8.0

The command is then dotnet-debugger-mcp, with no arguments. This needs ~/.dotnet/tools on your PATH, which is why it is not the default suggestion: a client that cannot find the command fails the same way a wrong path does.

Checking where a package came from

Releases after 0.6.1 carry a build provenance attestation: a statement, signed through GitHub, that this repository's release workflow built that package from a given commit. To check one before mirroring or installing it, download the .nupkg from nuget.org and run:

gh attestation verify DotnetDebugger.Mcp.<version>.nupkg --repo nevse/dotnet-debugger-mcp

nuget.org adds its own repository signature to every package it accepts, which changes the file, so the release workflow attests both: the package as built, attached to the GitHub Release, and nuget.org's copy, once it has checked that copy holds exactly what was built. For 0.6.2 only the first exists: verify the copy from gh release download v0.6.2 -R nevse/dotnet-debugger-mcp -p '*.nupkg'.

Upgrading

dotnet tool exec can keep running a version it has already downloaded, even after a newer one is published and indexed. Nothing looks wrong when it does: the server starts, the client reports it as connected, and the only sign is that the new release's tools are missing.

Measured while releasing 0.1.1. With only 0.1.0 in the local package folder, the unpinned command kept running 0.1.0, and clearing the NuGet HTTP cache did not change that. Naming the version once fetched the new package, after which the unpinned command used it too:

dotnet tool exec DotnetDebugger.Mcp --version 0.1.1 --yes

With the version pinned, upgrading is changing it in the command, and the question does not arise. An installed tool is upgraded explicitly too:

dotnet tool update -g DotnetDebugger.Mcp --version <new version>

To see which version is actually running, ask the client to list this server's tools, or read the first line the server writes to stderr — it names the version at startup. A client's own health check reports only that the process started, not what it is.

A worked example

Catching a program before it has run a single line, which is the case attaching cannot reach:

User: "My app throws before it prints anything. Find out why."
Claude: [launch_program("/path/to/bin/Debug/net10.0/MyApp.dll")]"Prepared, not running yet. Setting a breakpoint on the first line of Main."
Claude: [set_breakpoint("/path/to/Program.cs", 12)]"Breakpoint set. It is unverified for now — nothing can bind before the program hasloaded its modules — and takes effect when the program starts."
Claude: [start_program()]Claude: [wait_for_stop()]"Stopped at Program.cs:12, before a single line has run."
Claude: [step_over(thread_id: 1)]Claude: [get_variables(frame_id: 0)]"configPath is null, and the next line passes it to File.ReadAllText."
Claude: [get_program_output()]"The program printed nothing before the throw, which matches what you saw."

Tools

Each tool declares MCP annotations, so a client that reads them can approve the read-only ones without asking and still ask for the rest. The last column says which:

• read-only - reads the session or the machine and changes nothing.
• session - changes only the debug session: stepping, resuming, removing a breakpoint. Not destructive.
• runs getters - reads that run the debuggee's own code: the debugger calls ToString, property getters and DebuggerDisplay to show a value. Not marked read-only, though not destructive either.
• ends a launched program - destructive: a program the session launched is killed.
• runs code - destructive and open-world: it starts, builds or evaluates code the agent chose. A breakpoint's condition is a C# expression evaluated at every hit, so setting one counts.

The annotations are hints, and a client is free to ignore them.

Debugging

.NET MAUI

Needs a one-off setup - see Debugging a .NET MAUI app.

Breakpoints

Execution and inspection

Documentation

Things worth knowing

Breakpoints need portable PDBs

A breakpoint binds through the target's symbols, so the debuggee must be built with portable PDBs sitting next to its assembly. A missing or mismatched PDB is the most common reason set_breakpoint answers verified: false with No symbols have been loaded for this document.

Debug builds already do this. For a Release build, or any project that changes the defaults:

<PropertyGroup>  <DebugType>portable</DebugType>  <DebugSymbols>true</DebugSymbols></PropertyGroup>

Optimized code also moves locals out of reach, so get_variables is only fully useful with <Optimize>false</Optimize>.

macOS: headless environments may need an entitlement

macOS will not let a debugger take another process's task port unless the target carries com.apple.security.get-task-allow. A program run through the dotnet muxer is fine, because the muxer ships with that entitlement; an apphost produced by dotnet publish is ad-hoc signed with no entitlements at all.

The debugger's own side of this needs nothing from you: the debug adapter is started through the dotnet muxer as well, so it inherits the entitlements that let it debug at all.

On a desktop session this does not affect you. Debugging a self-contained publish carrying no entitlements is verified to work there.

It matters in a headless environment, such as a CI runner, and it fails badly when it does: macOS refuses by blocking rather than returning an error, so the debugger stops responding and the call never comes back. If a launch or attach hangs on macOS with no error at all, sign the target and try again:

codesign -s - -f --entitlements get-task-allow.entitlements ./MyApp

with get-task-allow.entitlements containing:

<?xml version="1.0" encoding="UTF-8"?><!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd"><plist version="1.0"><dict>  <key>com.apple.security.get-task-allow</key>  <true/></dict></plist>

What Just My Code changes about stepping

DOTNET_DEBUGGER_MCP_JUST_MY_CODE=false does not make a step stop in code you have no symbols for. Neither setting does: a step that lands in a module without symbols steps straight back out, because there is no source to report a location against. A step through an interpolated string goes through System.Private.CoreLib and comes back to the next statement of your own method either way.

What the setting changes is which modules get their symbols looked for at all. With Just My Code on, only assemblies built by you are searched; with it off, every module is, so a step does surface inside a dependency you happen to have symbols for. That is what to turn it off for.

What launch_program will run

The debugger starts whatever it is given, so launch_program only accepts a .NET app: a .dll with its .runtimeconfig.json beside it, or an apphost with the app's .dll and .runtimeconfig.json beside it under the same name. That covers a normal build and a self-contained publish. A shell, a system tool or any other native executable is refused, as is a single-file publish, which cannot be debugged anyway.

Environment variables that load other code into the program are refused too, since any of them turns launching a legitimate app into running something else: DOTNET_STARTUP_HOOKS, DOTNET_ADDITIONAL_DEPS, DOTNET_SHARED_STORE, DOTNET_ROOT*, the CoreCLR profiler variables (CORECLR_ENABLE_PROFILING, CORECLR_PROFILER*, COR_ENABLE_PROFILING, COR_PROFILER*), ASP.NET Core's hosting startup assemblies (ASPNETCORE_HOSTINGSTARTUPASSEMBLIES, DOTNET_HOSTINGSTARTUPASSEMBLIES), LD_PRELOAD, LD_LIBRARY_PATH, LD_AUDIT and every DYLD_*.

A launchSettings.json profile is read by this server, not handed to the debugger. Pass project_path (the project file or its directory) and optionally launch_profile. With no name given, the https profile is used, or else the first profile with commandName: "Project". Its commandLineArgs, workingDirectory and environmentVariables become the launch's arguments, working directory and environment, and applicationUrl becomes ASPNETCORE_URLS. These are then checked the same way as the parameters above. Arguments and a working directory passed to launch_program replace the profile's. Passed environment variables are added on top of the profile's. Only Project profiles are accepted, because an Executable profile names a program of its own, and executablePath is ignored, so the program is always the one passed in. Values are used literally: neither $(...) nor %VAR% is expanded. The profile comes with the repository, so it gets no more trust than the agent does: a startup hook set in it is refused like any other, and the error names the profile. A project with no launchSettings.json is launched without a profile unless one was named.

This is not a sandbox, and it does not stop launch_program from serving as a shell. A .NET app can do anything its user can, an agent able to write and build code can launch what it built, and ready-made .NET apps that run other programs are already on the machine: the SDK's own sdk/<version>/dotnet.dll has its .runtimeconfig.json beside it, so launching it reaches the whole dotnet CLI, building and running any project included. evaluate_expression can call Process.Start in whatever is being debugged. What the check stops is the agent naming a native program directly. Treat a client that allows this server's tools as one that allows a shell, and confirm launch_program, evaluate_expression, build_mobile_app and attach_to_process calls rather than approving them automatically, along with set_breakpoint and set_function_breakpoint, whose condition is code run at every hit. All of these are annotated as destructive and open-world, together with start_program and launch_mobile_app (see Tools), so a client that honours the annotations can be set to ask for those and approve the reads.

Secrets in what the debugger shows

Values read from the debuggee go to the model, so get_variables, expand_variable, evaluate_expression, get_exception_info and get_program_output mask what looks like a secret first, as [redacted: N chars]. Two rules decide it:

• By name. A string whose variable or member name contains password, passwd, pwd, secret, token, apiKey, connectionString, credential, privateKey, accessKey or authorization is masked whole. Only strings: an int tokenCount or a CancellationToken token shows as it is. An object is not masked, and its members are checked when they are expanded. For evaluate_expression the name is the last member the expression reads, so config.Db.Password counts.
• By format. In any text, named or not: Bearer and Basic credentials, JWTs, Password=, Pwd=, ApiKey=, sig= and the like inside connection strings and URLs, keys with a well-known prefix (sk-, ghp_, github_pat_, AKIA, xoxb-, AIza), and PEM private keys. Only the secret itself is masked, so the rest of a connection string stays readable.

null and an empty string are shown as they are, since whether a secret is set at all is often the question. Each response says how many values it masked in redacted.

The server's own log, which MCP clients usually keep on disk, records only how long each line of program output was. The text itself is logged at DOTNET_DEBUGGER_MCP_LOG_LEVEL=Debug and below, masked one line at a time, so a secret printed over several lines, such as a PEM key, may be masked only in part there. Errors the debugger reports after a request gave up waiting, and the debugger's own diagnostic output, are masked too.

This reduces exposure; it does not prevent it. A secret in a variable called s, in a format the patterns do not know, or read piecemeal - evaluate_expression of password.Length or password[0] - passes. Set DOTNET_DEBUGGER_MCP_REDACT_SECRETS=false to see values unmasked; the agent has no way to turn it off itself.

Running code in the program

evaluate_expression runs whatever the expression calls - a property getter, ToString, any method, Process.Start included - with the program's privileges, and listing variables runs property getters, ToString overrides and DebuggerTypeProxy constructors to show values. That is how a debugger shows what a program holds, and DOTNET_DEBUGGER_MCP_EVAL_TIMEOUT_MS is the only bound on it by default.

Set DOTNET_DEBUGGER_MCP_ALLOW_FUNC_EVAL=false to stop the debugger running the program's code, whether to show a value or because an expression asks it to. Fields, locals and arguments are read as usual. A property is listed as [not evaluated: function evaluation is off], an object shows as its type name in braces, and an expression that calls anything - a getter, a method, a constructor - fails. The agent has no way to turn it back on. Two things still run code:

• get_exception_info calls the exception's Message and InnerException getters to report them.
• Reading a static field of a type that has not been initialized yet runs that type's static constructor, in evaluate_expression and when static members are expanded alike.

A breakpoint condition is an expression too, so one that calls a getter or a method - items.Count > 3

• cannot be evaluated with this off. Such a breakpoint stops at every hit, as if it had no condition, and nothing in the stop says why. Compare fields instead, or leave the condition out and check after the stop.

This narrows what debugging can run; it does not make the server safe to hand to an agent you do not trust. launch_program still starts any .NET program.

Attaching to other users' processes

A debugger can read and change everything in the process it attaches to, so by default this server attaches only to processes belonging to the user it runs as. attach_to_process refuses anything else before it looks at the process at all, and list_dotnet_processes marks each entry with an owner of current_user, other_user or unknown.

unknown is refused as well. On Windows that is what a system or elevated process looks like, since its token cannot be opened, and treating it as your own would make the check decorative wherever the lookup does not work.

DOTNET_DEBUGGER_MCP_ALLOW_OTHER_USER_PROCESSES=true lifts the restriction. The operating system still has its own say: on Linux and macOS a normal user cannot attach to another user's process even with this enabled, so in practice it matters when the server runs elevated or as root.

Debugging more than one process

By default the server debugs one process at a time, and DOTNET_DEBUGGER_MCP_MAX_SESSIONS raises that. Each session has its own process, breakpoints and stops.

attach_to_process and launch_program return a session_id. While only one session is open you can ignore it, because every tool defaults to the only session. Taking on a second process opens a second session rather than failing, and from then on session_id becomes required: with two processes open, guessing which one a continue_execution was meant for would be worse than asking.

detach_from_process leaves the session open and free, so the next attach or launch reuses it rather than taking another slot.

The default of one is deliberate. Every attach carries a risk of a native crash inside the debugging shim, so more sessions means more exposure. What such a crash costs is bounded: each session drives its own debug adapter in a process of its own, so the one that crashes takes its session with it and leaves the server and any other session running.

Debugging a .NET MAUI app

A MAUI app is debugged the same way a desktop program is - prepare, set breakpoints, start - but two things have to be true before it can be, and neither is true of an app built the ordinary way.

It has to run on CoreCLR. This debugger attaches to CoreCLR and nothing else, while a MAUI app still builds against Mono by default. CoreCLR is there from net10.0-android, and from net11.0 for -ios and -maccatalyst; below those versions the app would build, start, and never connect back, so build_mobile_app refuses them and says which version would work.

It has to carry the remote debugging library. Debugging a runtime on a phone takes a native library inside the app, which opens the connection the debugger attaches through, and a matching one on this side. Both are open source, part of clrdbg, and ship in this package from the clrdbg release the debugger is built from, so there is nothing to install or set up:

remote-coreclr/  remote-host/     osx-arm64/, osx-x64/, win-x64/, win-arm64/, linux-x64/, linux-arm64/  remote-target/   android/, ios/, maccatalyst/

To use a copy built yourself instead - one with tracing compiled in, say - point DOTNET_DEBUGGER_MCP_REMOTE_CORECLR_HOST at a remote-host directory, DOTNET_DEBUGGER_MCP_REMOTE_CORECLR_TARGET at a remote-target one, or both. Either half left unset comes from the bundled copy.

Visual Studio's own remote debugger libraries (vsdbg*) are not used, wherever they come from: a directory holding a file with vsdbg in its name is refused, bundled or configured.

Then the session looks like this:

list_mobile_devices()                                → ids to choose frombuild_mobile_app(project_path, device_id)            → the .apk or .app, built to be debuggablelaunch_mobile_app(project_path, device_id)           → prepared, not startedset_breakpoint(file, line)                           → in place before the app's first linestart_program()                                      → boots, installs, launches: minutes, not secondswait_for_stop()                                      → and from here it is an ordinary debug session

build_mobile_app adds an MSBuild file of its own to the build - it ships beside the server as Resources/CopyRemoteCoreclrTargetLibrary.targets, and is readable - which puts the target library into the app bundle or the Android package. Nothing in your project changes.

Some notes on what each platform does:

• Android. Pick an emulator by its AVD name; it does not have to be running, because the debugger boots a cold one itself. A phone is picked by its adb serial. The app is built for every ABI at once, so there is no runtime identifier to choose. The app's logcat output arrives in get_program_output. uninstall_app: true clears a previous install first, which is what to reach for when a signature or a permission has changed.
• iOS. A simulator is picked by its UDID and booted if it is not running. A physical device has to be provisioned for the app the usual way - the debugger installs and launches it, it does not sign it.
• Mac Catalyst. The device id is maccatalyst; there is nothing to boot or install, which makes it the quickest way to check that the setup works at all.

process_id stays null throughout: the app runs on the device, so its pid is not one this machine could be pointed at. Everything else - breakpoints, stepping, locals, expressions, exceptions - behaves exactly as it does locally.

How failures are reported

Every tool reports a failure the same way:

{  "success": false,  "error": "Returned from a call to Continue that was not matched with a stopping event. (0x8013132F)",  "explanation": "The process was already running, so there was nothing to resume. Check get_process_status before continuing."}

error is whatever the debugger said, kept verbatim. explanation says what the failure means and what to do about it, for the CORDBG_E_* results the debugger raises: a process that has exited, an operation that needs the debuggee stopped, a variable that is not live at this instruction, a frame id from an earlier stop, another debugger already attached. It is null for failures that are not the debugger's, such as invalid arguments.

Configuration

Set these as environment variables in your client's configuration:

{  "mcpServers": {    "dotnet-debugger": {      "command": "dotnet",      "args": ["tool", "exec", "DotnetDebugger.Mcp", "--version", "0.8.0", "--yes"],      "env": {        "DOTNET_DEBUGGER_MCP_LOG_LEVEL": "Debug"      }    }  }}

Troubleshooting

The server does not appear in the client. Check that the path in the configuration is absolute and correct, then fully quit and restart the client. Claude Desktop logs to ~/Library/Logs/Claude/ on macOS and %APPDATA%\Claude\logs\ on Windows.

Attaching fails. The usual causes are that the process is owned by another user (see above), has already exited, is not a .NET process, or already has a debugger attached. list_dotnet_processes shows what the server can see and who owns it.

A breakpoint is not hit. Check that it came back verified: true. If not, the PDB is the first thing to look at. If it is verified and still not hit, confirm the line is actually reached and that the file path matches the one the PDB records.

list_dotnet_processes comes back empty on macOS or Linux. Module enumeration needs permissions there, and the server falls back to detection by process name, which misses programs whose apphost is renamed.

A launch or attach hangs on macOS with no error. See the entitlement section above.

A MAUI app starts and never stops at a breakpoint. It is almost certainly running on Mono rather than CoreCLR - which happens when it was built by something other than build_mobile_app, since an ordinary dotnet build produces neither the runtime nor the library the debugger needs. Rebuild with build_mobile_app and launch the app it reports.

list_mobile_devices shows no Android devices. It reads emulators from the AVD directory and connected phones from adb, so it needs the Android SDK: set ANDROID_HOME or ANDROID_SDK_ROOT if it is somewhere unusual. The warnings in the response say which source could not be read and why.

To see what the server is doing, set DOTNET_DEBUGGER_MCP_LOG_LEVEL=Trace and DOTNET_DEBUGGER_MCP_ENABLE_DIAGNOSTICS=true, or run it directly and watch stderr:

dotnet tool exec DotnetDebugger.Mcp --version 0.8.0 --yes

Development

git clone https://github.com/nevse/dotnet-debugger-mcp.gitcd dotnet-debugger-mcpdotnet builddotnet test

The build takes the debug adapter from the clrdbg release ClrdbgVersion names, checks its hash and puts it next to the server, so there is no separate step. To build the adapter from a clrdbg checkout instead - a fork carrying a fix, say - point ClrdbgSourcePath at it:

ClrdbgSourcePath=~/work/clrdbg dotnet build

The integration tests drive a real debuggee with real breakpoints, so they are slower than the rest and are separated by TestCategory=Integration.

To point a client at your working copy instead of the published package:

claude mcp add dotnet-debugger -- dotnet run --project "$(pwd)/src/DotnetDebugger.Mcp/DotnetDebugger.Mcp.csproj"

CONTRIBUTING.md covers the layout and conventions. docs/RELEASING.md covers cutting a release, including the nuget.org trusted-publishing policy that the workflow depends on but does not show.

Related projects

• clrdbg — the .NET debugger this server drives
• SharpDbg — the debugger it ran on before the move
• ClrDebug — ICorDebug API wrapper
• Model Context Protocol — the specification and SDKs

License

MIT