netcoredbg-mcp
Enables AI coding agents to debug .NET applications with breakpoints, stepping, variable inspection, and GUI automation for WPF, WinForms, and Avalonia apps.
README
netcoredbg-mcp
netcoredbg-mcp gives AI coding agents a real debugger for .NET applications.
Through the Model Context Protocol, an agent can launch or attach to a process,
set breakpoints, step through code, inspect variables, evaluate expressions, read
debug output, and operate Windows UI Automation surfaces such as WPF, WinForms,
and Avalonia windows without opening an IDE.
103 MCP tools · 8 prompts · 4 resources · 968 collected tests · release v0.14.0
Quick Links
- Start: Quick Start · Installation · Client Setup
- Use: First Debug Session · GUI App Debugging · Visual Inspection
- Reference: Available Tools · Resources · Prompts · Architecture
- Project: Contributing · Changelog · License
What's New in v0.14.0
- Runtime smoke v2 state oracle —
run_runtime_smokenow acceptsnetcoredbg.runtime_smoke.v2plans with baseline setup, case transitions, before/after probes, diffs, cleanup aggregation, and compact evidence. - Probe and template expansion — v2 adds UI property/text/grid probes, debug evaluation, tracepoints, output checks, JSONPath file assertions, process metrics, key-sequence actions, and generated A/B cases.
- Release-gated fixture coverage — critical tests and manual smoke entries now cover WPF and Avalonia state-oracle scenarios with runnable JSON examples.
- Adapter failure hardening — bridge and process-registry failures now
return structured
BLOCKEDevidence instead of crashing the smoke runner or producing false PASS results.
Highlights
| Capability | What agents can do |
|---|---|
| Debug control | Launch, attach, restart, continue, pause, terminate, and step through .NET code |
| Breakpoints | File, function, conditional, hit-count, exception, and tracepoint workflows |
| Inspection | Threads, stack frames, scopes, variables, modules, expressions, source, disassembly, memory |
| GUI automation | Window trees, element search, clicks, keystrokes, screenshots, annotations, clipboard, window management, UI evidence |
| Build integration | Pre-launch dotnet build, progress notifications, build diagnostics, cleanup of locked debug processes |
| Runtime smoke | Hygiene preflight, instrumentation groups, output checkpoints, freshness checks, bounded scenario runner |
| Multi-agent safety | Session ownership through mcp-mux, read-only observers, inactivity release |
Quick Start
# 1. Install the MCP server
pipx install netcoredbg-mcp
# 2. Run first-time setup
netcoredbg-mcp --setup
# 3. Register it in Claude Code
claude mcp add netcoredbg -- netcoredbg-mcp --project-from-cwd
Then ask your agent:
Set a breakpoint in Program.cs, run the app, and inspect local variables when it stops.
Critical Notes
[!IMPORTANT] For .NET Core debugging, the
dbgshim.dllnext tonetcoredbg.exemust match the target runtime major version. The setup wizard scans installed runtimes and prepares compatible dbgshim copies.
[!IMPORTANT]
start_debugis a long-poll tool. If the debuggee is a GUI app, it may return only after the app stops at a breakpoint, exits, or times out. Use screenshots and UI tools while the app is running; use variable inspection only when stopped.
[!CAUTION] Do not commit
.mcp.json,.netcoredbg-mcp.launch.json, credentials, server inventory, or local downstream project paths. Launch profiles supportinheritso secrets can stay in the MCP server process environment.
Installation
Requirements
- Windows for GUI automation and FlaUI/pywinauto workflows.
- Python 3.10 or newer.
- .NET SDK/runtime for the target application.
netcoredbg; usenetcoredbg-mcp --setupunless you need a custom install.- An MCP client such as Claude Code, Cursor, Cline, Roo Code, Windsurf, Continue, or Claude Desktop.
Recommended Install
pipx install netcoredbg-mcp
netcoredbg-mcp --setup
netcoredbg-mcp --version
The setup wizard downloads or discovers netcoredbg, scans dbgshim versions,
builds the FlaUI bridge when needed, and prints a ready-to-use MCP configuration
snippet.
Manual Install
pip install netcoredbg-mcp
$env:NETCOREDBG_PATH = "C:\Tools\netcoredbg\netcoredbg.exe"
netcoredbg-mcp --project-from-cwd
Use a manual install when you pin a locally managed netcoredbg build or when a
corporate environment blocks automatic downloads.
Upgrading
pipx upgrade netcoredbg-mcp
netcoredbg-mcp --setup
Run setup after an upgrade when the target .NET runtime changed, when you need a new FlaUI bridge build, or when MCP client snippets should be regenerated.
Configuration
Project Launch Profiles
start_debug can read .netcoredbg-mcp.launch.json from the resolved project
root and apply profile environment variables to the debuggee process. The build
process environment is not changed.
{
"defaultProfile": "default",
"profiles": {
"default": {
"env": {
"DOTNET_ENVIRONMENT": "Development",
"APP_MODE": "Debug"
},
"inherit": ["PATH"]
}
}
}
Precedence is deterministic:
inheritcopies only explicitly named variables from the MCP server process.- Profile
envvalues override inherited values. - Direct
start_debug(env={...})values override the profile.
env values set to null are passed through to DAP as explicit nulls to
request variable removal or unset semantics where the adapter supports it. Tool
responses include only variable names, counts, profile name, source path, and
redacted metadata; they never echo environment values.
The repository .gitignore excludes .netcoredbg-mcp.launch.json by default.
Commit a profile only when it contains non-secret, shareable values.
Base Server Configuration
Use --project-from-cwd for CLI-based agents that start the server from the
workspace. Use --project when the MCP client starts from a stable global
location and you want to constrain all debug paths explicitly.
{
"mcpServers": {
"netcoredbg": {
"command": "netcoredbg-mcp",
"args": ["--project-from-cwd"]
}
}
}
If setup did not install a managed netcoredbg, add NETCOREDBG_PATH:
{
"mcpServers": {
"netcoredbg": {
"command": "netcoredbg-mcp",
"args": ["--project-from-cwd"],
"env": {
"NETCOREDBG_PATH": "C:\\Tools\\netcoredbg\\netcoredbg.exe"
}
}
}
}
Project-Scoped Config
Use a project-local MCP config when a client supports it. Keep machine-specific secrets and binary paths outside git.
{
"mcpServers": {
"netcoredbg": {
"command": "netcoredbg-mcp",
"args": ["--project", "C:\\Work\\MyDotNetApp"]
}
}
}
Client Setup
Claude Code
claude mcp add netcoredbg -- netcoredbg-mcp --project-from-cwd
Cursor, Cline, Roo Code, Windsurf, Continue, Claude Desktop
Add the same server shape to the client-specific MCP configuration file:
{
"mcpServers": {
"netcoredbg": {
"command": "netcoredbg-mcp",
"args": ["--project-from-cwd"]
}
}
}
Common configuration locations:
| Client | Typical config path |
|---|---|
| Cursor | %USERPROFILE%\.cursor\mcp.json |
| Cline | VS Code extension MCP settings |
| Roo Code | %USERPROFILE%\.roo\mcp.json or project .roo\mcp.json |
| Windsurf | %USERPROFILE%\.codeium\windsurf\mcp_config.json |
| Continue | %USERPROFILE%\.continue\config.json |
| Claude Desktop | %APPDATA%\Claude\claude_desktop_config.json |
First Debug Session
The Long-Poll Pattern
Execution tools wait for a meaningful debugger event. start_debug,
continue_execution, step_over, step_into, and step_out return when the
debuggee stops, exits, terminates, or times out.
Typical Workflow
1. Add a breakpoint in the code path you want to inspect.
2. Start debugging with pre_build=true.
3. Wait for state=stopped.
4. Read call stack, scopes, and variables.
5. Evaluate focused expressions or step to the next line.
6. Continue or terminate the session.
Pre-Build Launch Example
{
"program": "bin/Debug/net8.0/MyApp.dll",
"build_project": "MyApp.csproj",
"pre_build": true,
"stop_at_entry": false
}
For .NET 6+ applications, passing a built .exe is accepted when a matching
.dll and .runtimeconfig.json exist. The server resolves the DLL target to
avoid deps.json conflicts.
GUI App Debugging
The Rule
Do not use debugger inspection tools while a GUI app is running normally. First stop at a breakpoint or pause the process; otherwise stack, scopes, and variables are unavailable by design.
GUI Workflow
1. Launch the app.
2. Use ui_take_screenshot or ui_get_window_tree to observe the running UI.
3. Use UI tools to click, type, select, or wait for state changes.
4. Set a breakpoint before the code path you need to inspect.
5. Trigger the UI action.
6. When state=stopped, inspect variables and call stack.
Stealth Mode
Use stealth mode when the user should keep focus in another application while the agent launches and inspects the debuggee in the background.
start_debug(
program="bin/Debug/net8.0/App.dll",
build_project="App.csproj",
stealth_mode=True,
)
ui_get_window_tree()
ui_click(automation_id="btnSave")
In stealth mode, UIA tree reads and automation-id clicks avoid activating the
debuggee window. ui_send_keys and blank-screenshot fallbacks may use
flash-focus: the bridge briefly brings the debuggee forward, performs the
operation, and restores the previous foreground window. Call
ui_bring_to_front() when the debuggee should explicitly leave stealth mode.
Visual Inspection
Screenshots return MCP image content, so vision-capable agents can inspect
layout and state. Annotated screenshots add Set-of-Mark labels for elements that
can be clicked with ui_click_annotated.
ui_take_screenshot()
ui_take_annotated_screenshot()
ui_click_annotated(element_id=3)
Runtime Smoke Evidence
For repeatable agent-side verification, use the runtime smoke tools together:
debug_hygiene_preflight clears stale debugger state, output_checkpoint and
output_assert_since prove output changed after a known point,
verify_debug_freshness checks that the live process still matches the expected
workspace and artifacts, and run_runtime_smoke executes a bounded scenario
plan with cleanup.
The manual smoke fixtures now cover the baseline console/WinForms app,
tests/fixtures/WpfSmokeApp, and tests/fixtures/AvaloniaSmokeApp. Build all
three fixture projects before claiming full GUI smoke coverage; missing fixture
binaries intentionally skip their corresponding manual scenarios.
For WPF product-smoke workflows, start from
docs/examples/runtime-smoke-wpf-workflow-plan.json.
It shows the accepted netcoredbg.runtime_smoke.v1 schema, steps operations
for DataGrid snapshots/assertions, scoped ListBox item actions, output
checkpoints, focus assertions, and cleanup.restore_files with graceful debug
stop. The example launches the WPF fixture DLL through dotnet, so freshness
expects expected_process_name: "dotnet" and expected_modules: ["WpfSmokeApp.dll"]. The matching manual scenario is
WPF One-Call Runtime Smoke Workflow.
The WPF workflow now connects UI automation eagerly after launch, chooses a
stable usable top-level window, merges structured GridPattern cell evidence
with descendant text fallback, and restores fixture files even when Windows
briefly holds attributes or locks. Avalonia remains a first-class compatibility
target: its manual fixture is expected to produce bounded UNSUPPORTED or
BLOCKED evidence for UIA gaps instead of being omitted from release checks.
For WPF DataGrid drag/drop and edge-scroll release proof, start from
docs/examples/runtime-smoke-v2-drag-drop-grid.json.
The v2 example shows ui.drag, before/after ui.grid.viewport probes,
selected payload identity checks, negative no-op expectations, and fail-closed
BLOCKED behavior when real pointer route or viewport evidence is unavailable.
WinForms dragList primitive smoke is not a substitute for WPF DataGrid CR-001
acceptance.
Edit-and-Continue
apply_code_change applies supported source edits to a stopped .NET debug
session without restarting the process. It is intended for method-body changes
found during a live investigation; it is not a replacement for rebuilding when
the shape of the program changes.
EnC Setup
Build an EnC-capable netcoredbg with the bundled setup flow:
netcoredbg-mcp setup --enc
The command runs scripts/build-netcoredbg-enc.ps1, installs the maintained
thebtf/netcoredbg fork build with ncdbhook.dll, and places it in the
managed ~/.netcoredbg-mcp/netcoredbg path that accepts the custom DAP
applyDeltas request. On Windows x64, setup downloads the published
3.1.3-1062-enc.2 EnC build by default; pass -BuildFromSource to the script
for a local source build. If ncdbhook.dll is missing, apply_code_change
returns an actionable error instead of crashing.
Runtime Code Change Workflow
find_code_symbol(name="OrderService")
get_source_context(file="Services/OrderService.cs", line=42, radius=8)
apply_code_change(
file="Services/OrderService.cs",
edits=[{"start_line": 42, "end_line": 44, "new_text": "return fixedValue;"}],
)
continue_execution()
The debug session must be in STOPPED state. Successful changes update the source
file and apply IL/metadata/PDB deltas through netcoredbg; the session remains
STOPPED until you continue or step.
Rude edits such as adding fields, changing method signatures, or changing
generics are rejected before runtime application. Use
restart_debug(rebuild=True) for those changes.
Available Tools
| Category | Count | Tools |
|---|---|---|
| Debug control | 12 | start_debug, attach_debug, stop_debug, restart_debug, continue_execution, pause_execution, step_over, get_step_in_targets, step_into, step_out, get_debug_state, terminate_debug |
| Breakpoints and exceptions | 6 | add_breakpoint, remove_breakpoint, list_breakpoints, clear_breakpoints, add_function_breakpoint, configure_exceptions |
| Inspection and DAP coverage | 15 | get_threads, get_call_stack, get_scopes, get_variables, evaluate_expression, set_variable, get_exception_info, get_modules, get_progress, get_loaded_sources, disassemble, get_locations, quick_evaluate, get_exception_context, get_stop_context |
| Tracepoints | 4 | add_tracepoint, remove_tracepoint, get_trace_log, clear_trace_log |
| Snapshots and object analysis | 5 | create_snapshot, diff_snapshots, list_snapshots, analyze_collection, summarize_object |
| Memory | 2 | read_memory, write_memory |
| Output and build diagnostics | 4 | get_output, search_output, get_output_tail, get_build_diagnostics |
| Runtime smoke orchestration | 8 | debug_hygiene_preflight, instrumentation_group_create, instrumentation_group_inspect, instrumentation_group_clear, output_checkpoint, output_assert_since, verify_debug_freshness, run_runtime_smoke |
| UI automation | 46 | Window tree, element search, focus, keyboard, mouse, screenshots, annotations, selection, clipboard, window management, expand/collapse, value setting, virtualization, grid evidence, UI snapshots, UI events |
| Code search | 4 | find_code_symbol, find_code_references, get_source_context, search_source |
| Edit-and-Continue | 1 | apply_code_change |
| Process management | 1 | cleanup_processes |
MCP Resources
| URI | Contents |
|---|---|
debug://state |
Current debug session state |
debug://breakpoints |
Active breakpoints and verification state |
debug://output |
Buffered debuggee and build output |
debug://threads |
Current thread list |
MCP Prompts
| Prompt | Use it for |
|---|---|
debug |
General debugging workflow guidance |
debug-gui |
WPF, WinForms, Avalonia, and UI Automation debugging |
debug-exception |
Exception-first investigation |
debug-visual |
Screenshot and Set-of-Mark workflows |
debug-mistakes |
Common agent debugging mistakes and recovery |
investigate |
Parameterized symptom investigation |
debug-scenario |
Scenario-specific debugging plans |
dap-escape-hatch |
Advanced DAP commands without first-class MCP wrappers |
Multi-Agent Safety
When served through mcp-mux, mutating debug tools are session-owned. One agent
can control a live debug session while other agents keep read-only observability
through state, output, screenshot, and inspection tools. Ownership auto-releases
after the configured inactivity timeout.
Architecture Overview
graph TB
subgraph MCP["MCP Server"]
MAIN["__main__.py"]
SERVER["server.py"]
PROMPTS["prompts.py"]
TOOLS["tools/*"]
SESSION["session/*"]
BUILD["build/*"]
UI["ui/*"]
SETUP["setup/*"]
end
subgraph DAP["Debug Adapter Protocol"]
CLIENT["dap/client.py"]
PROTOCOL["dap/protocol.py"]
EVENTS["dap/events.py"]
end
MAIN --> SERVER
SERVER --> PROMPTS
SERVER --> TOOLS
TOOLS --> SESSION
TOOLS --> BUILD
TOOLS --> UI
SESSION --> CLIENT
CLIENT --> PROTOCOL
CLIENT --> EVENTS
CLIENT <-->|stdio JSON-RPC| NETCOREDBG["netcoredbg"]
NETCOREDBG --> APP[".NET debuggee"]
SETUP --> NETCOREDBG
How It Works
__main__.pyparses CLI flags, configures project-root policy, and starts the FastMCP stdio server.server.pyregisters tools, prompts, resources, progress notifications, and session ownership checks.- Tool modules keep debugger control, breakpoints, inspection, memory, output, process cleanup, and UI automation separate.
SessionManagerowns debugger state, path validation, event handling, snapshots, tracepoints, output buffers, and process cleanup.DAPClientspeaks JSON-RPC over stdio tonetcoredbg.- UI automation uses a FlaUI bridge on Windows, with pywinauto fallback where supported.
Command Line Options
netcoredbg-mcp --help
netcoredbg-mcp --version
netcoredbg-mcp --setup
netcoredbg-mcp setup --enc
netcoredbg-mcp --project C:\Work\MyApp
netcoredbg-mcp --project-from-cwd
| Option | Purpose |
|---|---|
--version |
Print the package version |
--setup |
Run first-time setup and exit |
--project PATH |
Constrain debug operations to a specific project root |
setup --enc |
Build and install an EnC-capable netcoredbg with ncdbhook.dll |
--project-from-cwd |
Detect the project root from the process working directory and MCP roots |
--project and --project-from-cwd are mutually exclusive.
Environment Variables
| Variable | Default | Purpose |
|---|---|---|
NETCOREDBG_PATH |
auto-discovered after setup | Explicit path to netcoredbg |
NETCOREDBG_PROJECT_ROOT |
unset | Project root fallback |
MCP_PROJECT_ROOT |
unset | Generic MCP project root fallback |
NETCOREDBG_ALLOWED_PATHS |
empty | Additional comma-separated allowed path prefixes |
NETCOREDBG_SCREENSHOT_MAX_WIDTH |
1568 |
Max inline screenshot width |
NETCOREDBG_SCREENSHOT_QUALITY |
80 |
Screenshot compression quality |
NETCOREDBG_MAX_TRACE_ENTRIES |
1000 |
Tracepoint log capacity |
NETCOREDBG_EVALUATE_TIMEOUT |
0.5 |
Tracepoint expression timeout in seconds |
NETCOREDBG_RATE_LIMIT_INTERVAL |
0.1 |
Tracepoint hit rate-limit interval |
NETCOREDBG_MAX_SNAPSHOTS |
20 |
Snapshot capacity |
NETCOREDBG_MAX_VARS_PER_SNAPSHOT |
200 |
Variables captured per snapshot |
NETCOREDBG_MAX_OUTPUT_BYTES |
10000000 |
Total output buffer cap |
NETCOREDBG_MAX_OUTPUT_ENTRY |
100000 |
Single output entry cap |
NETCOREDBG_SESSION_TIMEOUT |
60.0 |
Multi-agent ownership inactivity timeout |
NETCOREDBG_STACKTRACE_DELAY_MS |
0 |
Diagnostic delay before stackTrace requests |
FLAUI_BRIDGE_PATH |
auto-discovered | Explicit FlaUI bridge executable path |
LOG_LEVEL |
INFO |
Python logging level |
LOG_FILE |
unset | Optional diagnostic log file |
Troubleshooting
netcoredbg is not found
Symptom: startup or start_debug reports that netcoredbg cannot be found.
Cause: setup has not installed a managed debugger and NETCOREDBG_PATH is
not set.
Fix: run netcoredbg-mcp --setup, or set NETCOREDBG_PATH to the explicit
netcoredbg.exe path.
Verify: netcoredbg-mcp --version succeeds and your MCP client can list the
server tools.
Breakpoints do not bind
Symptom: a breakpoint stays unverified or the process never stops where expected.
Cause: stale build output, wrong target DLL, optimized Release binaries, or a line without executable IL.
Fix: run with pre_build=True, debug the Debug configuration, confirm the
source file matches the built assembly, and inspect list_breakpoints() for
DAP-adjusted lines.
Verify: the breakpoint response reports verified=true or includes the DAP
line adjustment.
GUI appears frozen
Symptom: a WPF, WinForms, or Avalonia window stops repainting after a debug command.
Cause: the debugger stopped the UI thread at a breakpoint or pause.
Fix: inspect variables while stopped, then call continue_execution() before
expecting the GUI to respond to clicks or keystrokes.
Verify: get_debug_state() reports running and screenshots update again.
Path is rejected in a worktree
Symptom: launch or build fails with a path validation error.
Cause: the project root was resolved to a different checkout, or the worktree path is outside the allowed root set.
Fix: use --project-from-cwd from the active worktree, or add the worktree
prefix to NETCOREDBG_ALLOWED_PATHS.
Verify: start_debug accepts build and program paths under the worktree.
Limitations
- GUI automation is Windows-focused.
netcoredbgand DAP capabilities vary by runtime and target application.- Memory reads and writes require debug adapter support and valid memory references.
- Native debugging, browser automation, and non-.NET runtimes are out of scope.
Contributing
See CONTRIBUTING.md for setup, tests, PR expectations, and sensitive-data rules.
License
MIT. See LICENSE.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.