Windows Developer Agent MCP
Provides Git integration for inspecting working-tree status, diff statistics, and commit log pagination.
Integrates with the OpenAI Secure MCP Tunnel to establish an outbound encrypted tunnel to the OpenAI control plane, enabling secure remote access to the local MCP server.
Integrates pytest to run tests and return compact validation summaries.
Integrates Ruff to run linting and return compact validation summaries.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Windows Developer Agent MCPrefactor the process_order function and run the tests"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Windows Developer Agent MCP
A high-performance, full-access local Windows developer-agent backend powered by the Model Context Protocol (MCP Python SDK v2). The server provides atomic file operations, validated AST refactoring, durable background jobs, system diagnostics, and seamless integration with the OpenAI Secure MCP Tunnel.
Features
Robust Filesystem Tools: Atomic file creation, writing, copying, moving, and deletion; paginated listing and regex file search; complete reads with streaming byte cursors.
Precision Code Editing: Exact match, anchored replacements, and AST function/class body substitutions with syntax validation and rollback on syntax error.
Process & Command Execution: Synchronous and background execution for PowerShell, CMD, and native Windows executables with streaming spooling (1 MiB RAM limit rolling over to temp disk).
Durable Job Store: SQLite-backed background job queue (
.agent_state/jobs.sqlite3) with idempotency keys, progress tracking, and survivor workers surviving supervisor restarts.System Diagnostics: Live inspection of CPU, memory, disks, environment variables, installed applications, and Windows services.
Codebase Intelligence: Fast Python AST parsing for classes, functions, imports, and project summaries.
Validation Suite: Compact summaries from integrated
pytest,Ruff, andmypytools.Git Integration: Working-tree status, diff statistics, and commit log pagination.
Optional Browser Automation: Playwright browser automation with screenshot capture and UI interaction.
Resilient Supervisor & Control Panel: Heartbeat watchdog, automatic backoff recovery, and a loopback-only control panel at
http://127.0.0.1:8766.
Related MCP server: winfolio
Architecture
The application runs in two primary modes:
Secure MCP Tunnel Mode (Default):
START_MCP.batlaunchesscripts/bootstrap.ps1.Bootstraps
.venv(Python 3.10+), installs requirements, and executes startup smoke checks.Launches
scripts/supervisor.pymanagingtunnel-client.exe.tunnel-client.execreates an outbound encrypted tunnel to the OpenAI control plane, invoking companioncloudflared.exefor Cloudflare edge routing.Outbound requests from the control plane are forwarded over loopback to the local MCP server (
main.py).
Local HTTP Mode:
Bypasses external tunnels entirely.
Starts a Streamable HTTP MCP server listening on
http://127.0.0.1:8765/mcp.
Requirements
Operating System: Windows 10 / Windows 11 (x86-64 / amd64)
Python: Python 3.10 or newer (managed automatically in
.venvby the launcher)PowerShell: Windows PowerShell 5.1 or PowerShell 7+
Credentials (Tunnel Mode only): A valid control-plane API key stored locally outside version control.
Quick Start
Starting in Default Tunnel Mode
Double-click
START_MCP.bator run:START_MCP.batKeep the launcher window open. The launcher runs dependency checks, smoke validation, and starts the runtime supervisor.
Access the local operator panel at
http://127.0.0.1:8766/.
To stop the service, press Ctrl+C in the terminal window or click Stop in the local control panel.
Local HTTP Mode
To run locally without connecting to the OpenAI Tunnel:
$env:MCP_START_MODE = 'local-http'
.\START_MCP.batThe local Streamable HTTP MCP endpoint will be available at:
http://127.0.0.1:8765/mcp
Local Configuration
Control Plane API Key
When operating in Tunnel mode, the tunnel daemon requires a control-plane API key. This key is loaded locally using one of the following mechanisms:
Secret File (Recommended): Place the API key in:
.secrets\control_plane_api_key.txt(This path is strictly ignored by
.gitignoreand is never committed).Environment Variable: Set
CONTROL_PLANE_API_KEYbefore starting:$env:CONTROL_PLANE_API_KEY = "your-api-key"
Tunnel Profile
By default, the launcher automatically detects any configured local tunnel profile. To specify an explicit profile:
$env:MCP_TUNNEL_PROFILE = "<your-profile>"
.\START_MCP.batBundled Tunnel Runtime
This repository includes pre-built third-party runtime binaries required for the OpenAI Secure MCP Tunnel:
tunnel-client.exe: OpenAI Secure MCP Tunnel client daemon.cloudflared.exe: Cloudflare Tunnel companion daemon.cloudflared-manifest.json: Upstream pinned build manifest.
Why Both Executables Are Included
The project utilizes OpenAI's Tunnel client (openai/tunnel-client). The supported Windows distribution model places tunnel-client.exe and cloudflared.exe side by side. When establishing a tunnel over Cloudflare edge transport, tunnel-client.exe internally discovers and executes cloudflared.exe from its adjacent directory.
Even though this project's Python codebase does not invoke cloudflared.exe directly, deleting cloudflared.exe will break tunnel connectivity. Both executables and their manifest must remain together as a matched set.
For detailed provenance, versions, and cryptographic hashes, see BINARY_PROVENANCE.md.
Updating the Bundled Tunnel Runtime
When an upstream release of the tunnel client is available, maintainers must treat the runtime components as a matched set (tunnel-client.exe, cloudflared.exe, cloudflared-manifest.json). Users should not independently replace only one component unless official upstream compatibility documentation explicitly permits it.
Follow this verified 19-step update procedure:
Official Upstream Releases Only: Go ONLY to the official OpenAI repository releases: https://github.com/openai/tunnel-client/releases
Select Architecture: Select the correct official Windows architecture/build (e.g.,
windows-amd64).Prefer Bundled Release Artifact: Prefer the official bundled release artifact containing the compatible
tunnel-client+cloudflaredpair and associated manifest/provenance files (e.g.,tunnel-client-v<version>-windows-amd64.zip).Download Integrity Metadata: Download integrity metadata supplied by the release, such as
SHA256SUMS.txt, provenance/signature files, or equivalent, when provided upstream.Verify Downloaded Artifact: Verify the downloaded artifact hash BEFORE replacing repository binaries:
Get-FileHash -Algorithm SHA256 .\tunnel-client-v<version>-windows-amd64.zipNo Unofficial Mirrors: Never download replacement executables from mirrors, forums, random file hosts, or unofficial repositories.
Extract Outside Project: Extract the official bundle into a temporary directory outside the project first.
Confirm Versions: Confirm component versions before replacement:
.\tunnel-client.exe --version .\cloudflared.exe --versionReplace Matched Set Together: Replace the intended matched runtime files together in the repository root (
tunnel-client.exe,cloudflared.exe, andcloudflared-manifest.json).Recompute Local Checksums: Recompute local SHA-256 checksums after replacement:
Get-FileHash -Algorithm SHA256 .\tunnel-client.exe, .\cloudflared.exe, .\cloudflared-manifest.jsonUpdate Provenance Documentation: Update BINARY_PROVENANCE.md with:
New
tunnel-client.exeversion, commit SHA, and hash.New
cloudflared.exeversion, build date, and hash.New
cloudflared-manifest.jsonhash.Upstream release tag, commit, and update date.
Update Third-Party Licensing: Update applicable third-party license/NOTICE files (THIRD_PARTY_NOTICES.md,
third_party/) if upstream licensing material changed.Run Safe Validation Suite: Execute the safe local validation suite:
.\.venv\Scripts\python.exe -m pytest .\.venv\Scripts\python.exe -m ruff check . .\.venv\Scripts\python.exe -m mypy core tools scripts tests .\.venv\Scripts\python.exe -m compileall -q core tools scripts testsLocal HTTP Smoke Test: Perform a local HTTP smoke test that does not require production credentials:
$env:MCP_START_MODE = 'local-http' powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\bootstrap.ps1Local Tunnel Validation: Perform Tunnel validation separately using the maintainer's LOCAL credential/profile stored outside version control.
Run Tunnel Doctor: Run the Tunnel doctor command using a generic profile placeholder:
.\tunnel-client.exe doctor --profile <your-profile>Confirm Startup: Confirm startup still works cleanly with
START_MCP.bat.Review Changes: Review git status, working tree diffs, binary hashes, BINARY_PROVENANCE.md, and THIRD_PARTY_NOTICES.md:
git status git diff BINARY_PROVENANCE.md THIRD_PARTY_NOTICES.mdCommit Jointly: Commit the updated binary set and provenance documentation together in a single commit.
Local Control Panel
The supervisor exposes a local loopback management interface at http://127.0.0.1:8766/:
Health Monitoring: Displays live health state of the MCP server, heartbeat timestamp, and Tunnel
/readyzendpoint.Process Management: Shows active child processes, memory usage, and allows clean Restart or Stop.
Log Streaming: Live view of rotated supervisor events with automatic redaction of sensitive credentials.
Security Boundary: The panel binds strictly to loopback (
127.0.0.1), enforces strict Host checking, validates browser Origin headers, and requires a dynamic per-session CSRF token (X-Control-Token).
Testing and Validation
The project includes an automated test suite, static type checking, and linter validation:
# Run unit and integration tests
.\.venv\Scripts\python.exe -m pytest
# Run Ruff linter
.\.venv\Scripts\python.exe -m ruff check .
# Run Mypy static type checker
.\.venv\Scripts\python.exe -m mypy core tools scripts tests
# Run MCP startup smoke check
.\.venv\Scripts\python.exe -m scripts.health_check --jsonSecurity Notes
Credentials Outside Version Control: The repository strictly ignores
.secrets/and.env*. Never stage or commit credentials.Log Masking: The runtime supervisor and launcher automatically mask
CONTROL_PLANE_API_KEYwith[redacted]before writing to logs.Loopback Enforcement: All internal HTTP servers (MCP Streamable HTTP on port
8765, Control Panel on port8766) bind exclusively to127.0.0.1.Local State Isolation: Working databases (
.agent_state/jobs.sqlite3), session caches, and generated artifacts are stored in.agent_state/and ignored by Git.Audit Logging: Destructive and mutating file operations log metadata (timestamps, tool name, targets) to
.agent_state/audit.jsonl. Content bodies and secret arguments are excluded from audit records.
Project Structure
.
├── .gitignore # Enterprise-grade Git ignore configuration
├── .gitattributes # Binary and text line-ending rules
├── .editorconfig # Formatting and encoding standards
├── BINARY_PROVENANCE.md # Provenance and SHA-256 records for bundled binaries
├── CONTRIBUTING.md # Development and contribution guide
├── LICENSE # Apache License 2.0
├── NOTICE # Project copyright and third-party attribution
├── README.md # This documentation
├── SECURITY.md # Vulnerability reporting guidelines
├── START_MCP.bat # User-facing Windows launcher
├── THIRD_PARTY_NOTICES.md # Attribution for bundled third-party code
├── cloudflared-manifest.json # Upstream Cloudflare build manifest
├── cloudflared.exe # Bundled Cloudflare companion binary
├── pyproject.toml # Tool configuration (pytest, ruff, mypy)
├── requirements.txt # Core Python dependencies
├── requirements-browser.txt # Optional browser automation dependencies
├── tunnel-client.exe # Bundled OpenAI Secure MCP Tunnel client
├── core/ # MCP server core engine, executor, jobs, config
├── memory/ # Project memory subsystem documentation
├── scripts/ # Bootstrap, health check, and supervisor
├── tests/ # Test suite covering tools, launcher, supervisor
├── third_party/ # Vendored upstream third-party licenses
└── tools/ # MCP tool implementations (fs, edit, win, git, etc.)Third-Party Components
This repository redistributes precompiled binary software components under the Apache License 2.0:
OpenAI Secure MCP Tunnel Client (
tunnel-client.exe), Copyright 2026 OpenAI.Cloudflare Tunnel Client (
cloudflared.exe), Copyright Cloudflare, Inc.
Full licensing details and upstream notices are available in THIRD_PARTY_NOTICES.md and BINARY_PROVENANCE.md.
License
This project is licensed under the Apache License, Version 2.0. See LICENSE and NOTICE for details.
Author
Ali Khalili
GitHub: https://github.com/dibbed/
This server cannot be deployed
Maintenance
Related MCP Connectors
Agent-native notes, tasks, dev-docs, vaults, sync & handoffs. MCP + OpenAPI dual surface.
Personal assistant MCP server with search, execute, packages, jobs, secrets, and integrations.
Nifty's MCP server — exposes tasks, projects, messages, and files as tools for AI agents.
MCP server for secureFlows: token-free URL builders and integration-linting tools for AI agents.
Related MCP Servers
- AlicenseNot gradedqualityBmaintenanceEnables AI clients to control local Windows development tools by exposing project files, code search, file editing, test execution, Git operations, and resource viewing through a secure MCP interface with permission controls.10Apache 2.0
- AlicenseNot gradedqualityBmaintenanceEnables AI assistants to perform comprehensive Windows system health diagnostics, security checks, developer workflows, and maintenance tasks via 85 tools with actionable recommendations, over MCP and OpenAI-compatible HTTP.MIT
- AlicenseNot gradedqualityAmaintenanceEnables AI clients to securely control and interact with a local Windows machine through 218 configurable tools for files, Git, processes, Windows UI, browser automation, WSL, Office, recovery, skills, and child MCP servers.4MIT
- AlicenseNot gradedqualityBmaintenanceProvides a Windows-first local AI-agent gateway with configurable tools for files, Git, processes, Windows automation, WSL, browser control, durable agent runs, memory, verification, and intelligent routing, while exposing a secure MCP endpoint for ChatGPT Web and a local web UI.9MIT