mcp-windbg is a Model Context Protocol server that lets AI models control WinDbg, the Windows debugger used to inspect crash dumps and debug user-mode or kernel software. Developers and investigators use it to run debugger commands, examine failures, and manage multiple live debugging sessions through natural-language requests. The catalogue add-ons connect coding agents to this debugging workflow.
Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add svnscha/mcp-windbg --skill debug-remotegit clone --depth 1 https://github.com/svnscha/mcp-windbgWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/svnscha/mcp-windbg/debug-remote)<a href="https://agentmods.dev/skills/svnscha/mcp-windbg/debug-remote"><img src="https://agentmods.dev/badge/skills/svnscha/mcp-windbg/debug-remote/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/svnscha/mcp-windbg/debug-remote"><img src="https://agentmods.dev/badge/skills/svnscha/mcp-windbg/debug-remote.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector pass
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00049 | $0.00921 |
| Opus 5.5 | $0.00020 | $0.00368 |
| Sonnet 5.5 | $0.00010 | $0.00184 |
| Haiku 4.5 | $0.00005 | $0.00092 |
Grade A, and why
debug-remote scanned grade A with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Runs shell commandslowCapability
Expected in a hook, worth knowing in a rule or an instructions file.
`subprocess.DEVNULL`), or a stdin pipe that reaches EOF, can exhaust Windows nonpaged How it starts
The opening of the file, as written. The whole thing — 92 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Debug a live user-mode process
MCP connection
Use the existing mcp-windbg MCP connection, whether it is launched by a plugin, a native executable, Python, or an HTTP service. Tool names below are base names; resolve them against the tools exposed by that connection rather than assuming a plugin-specific prefix. Keep each session on the server that opened it. If multiple servers match, use the user's selected server or ask which one. If the required tools are unavailable, report the missing connection or tool and help check its configuration; do not register a second server.
Attach to an existing WinDbg/CDB debug server with the mcp-windbg tools.
This is user-mode only - for a kernel target use the available kernel-debug skill.
The other end
Someone must already be hosting the target. In WinDbg or CDB on that machine:
.server tcp:port=5005
If they have not, say so rather than guessing a connection string - there is nothing to attach to yet.
Safety before starting a server
CDB 10.0.29661.1004 with stdin redirected to NUL (/dev/null or
subprocess.DEVNULL), or a stdin pipe that reaches EOF, can exhaust Windows nonpaged
pool and freeze the host. For unattended startup use:
cdb -server tcp:port=5005 -noio <program>
Keep -server and its transport first, followed by -noio. This disables local
console I/O; use the remote client for commands and output. Apply -noio only to
the external -server process, never to the MCP-managed -remote client. Never use
NUL/EOF stdin without -noio. Arrange cleanup before launching, including failed
connections and interrupted investigations; record the spawned server PID. If
Could not write to pipe, 1450 repeats, stop only your owned server instead of
retrying. See the CDB upstream report.
Connecting
open_cdb_remote with the connection string:
- TCP:
tcp:Port=5005,Server=hostname - Named pipe:
npipe:Pipe=pipename,Server=hostname
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday Changed · +22 lines e8bc40a2c430
- 1mo ago Changed · +10 lines 2e4cf9f3bf93
- 1mo ago First seen · 60 lines · 49 tokens per session scan A a508d635f522
debug-remote is a skill published in the GitHub repository svnscha/mcp-windbg (1,632 stars, last pushed yesterday), licensed MIT. It adds 49 tokens to every session and 921 once invoked, about $0.0002 per session on Opus 5.5. A static security scan graded it A with 1 finding (runs shell commands). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
post-mortem
Diagnose instruction defects and optionally submit Rosetta GitHub issue.
memorix-troubleshooting
Use when Memorix MCP, setup, project binding, HTTP control plane, hooks, skills, or agent integration is missing, stale, or failing.
qa-knowledge
To run QA engineering — requirements/gap analysis, scenario & spec design, test implementation, failure triage — over the QA knowledge base.
solr-query
To build and debug Solr queries: eDisMax, block join, JSON facets, kNN, explain.
debugging
To investigate errors, test failures, and unexpected behavior — root cause before fix.
self-learning
MUST activate on: execution failure/error, mistake, wrong/unexpected result, expected≠actual mismatch, 2 consecutive mismatches, unhappy/upset user, user asks why something failed/didn't work.