McpServerDotNet

McpServerDotNet is a .NET 10 Model Context Protocol server for local developer automation. It exposes a workspace-aware filesystem surface, MCP resources and prompts, an activity-routing layer for Codex-style workflows, and optional execution, SSH, web, and Ollama-backed local inference tooling.

This is not a generic scaffold anymore. The repository already contains a working stdio MCP host, real tool/resource/prompt routing, safety gates around higher-risk operations, optional local Ollama integration, and setup scripts for clients such as LM Studio and VS Code.

Who this is for

• Developers building or operating local MCP workflows from .NET.
• LM Studio or GitHub Copilot users who want a repo-aware local MCP server.
• Teams that want a Clean Architecture MCP host they can extend with their own tools, resources, prompts, and infrastructure adapters.

Implemented capabilities

Always available

• Stdio MCP host with JSON-RPC lifecycle handling for initialize, ping, shutdown, exit, tools, resources, and prompts.
• Workspace-aware filesystem tools for read, write, append, copy, move, delete, metadata, and directory listing.
• Workspace tools for root switching, project-folder switching, status, and bounded code-review snapshots.
• MCP resources for file text, directory listings, metadata, recursive tree snapshots, and recent workspace changes.
• MCP prompts for file summarization and grounded directory review.
• Activity tools that classify requests and build bounded local-model context packets without calling a model directly.
• Clean Architecture split across Host, Protocol, Application, Infrastructure, Contracts, and Domain projects.

Conditionally enabled

• shell.exec for non-interactive local command execution under an explicit allow/deny policy.
• inference.local_* tools for Ollama-backed summarization, planning, code review, and completion.
• ssh.execute and ssh.write_text for profile-based remote automation.
• web.fetch_url and web.search for allowlisted outbound web access.

Client integration

• LM Studio installer and diagnostics scripts.
• VS Code user-level MCP config installer.
• Packable dotnet tool host via McpServer.Host.
• Optional Ollama-backed local inference path for summarization, planning, code review, and bounded completion.

Solution layout

| Project | Responsibility | | --- | --- | | src/McpServer.Host | Host bootstrap, configuration, Autofac registrations, stdio loop, logging. | | src/McpServer.Protocol | MCP lifecycle, session state, JSON-RPC dispatch, tool/resource/prompt routing. | | src/McpServer.Application | Tool, resource, prompt, activity, and application abstractions. | | src/McpServer.Infrastructure | Filesystem, path policy, process execution, Ollama, SSH, web, logging implementations. | | src/McpServer.Contracts | MCP and JSON-RPC DTOs. | | src/McpServer.Domain | Reserved domain layer for future business rules. | | tests/* | Unit and stdio integration coverage. |

Default tool surface

Filesystem

• fs.write_text
• fs.append_text
• fs.read_file
• fs.read_text
• fs.get_metadata
• fs.list_directory
• fs.create_directory
• fs.move_path
• fs.copy_path
• fs.delete_path

Workspace and review

• workspace.set_root
• workspace.open
• workspace.select_folder
• workspace.status
• workspace.inspect

Activity orchestration

• activity.route
• activity.schemas.list
• activity.context.preview
• activity.run

Resources

• file:///workspace/...
• dir:///workspace/...
• filemeta:///workspace/...
• tree:///project
• tree:///workspace
• changes:///project
• changes:///workspace

Prompts

• prompt.summarize_file
• prompt.review_directory

Safety model

Higher-risk capabilities are off by default.

• Shell execution is disabled unless explicitly enabled and constrained by command policy.
• SSH tooling is disabled unless named profiles are configured.
• Web access is disabled unless enabled with an allowlist.
• Workspace paths are normalized and validated against allowed roots before file access.
• Destructive file operations go through dedicated policy checks and per-path locking.
• Logs are written to logs/ so stdio MCP traffic stays clean.

See SECURITY.md and docs/workspace-configuration.md for the operational details.

Run locally

Restore and build the solution:

dotnet restore .\McpServer.slnx
dotnet build .\McpServer.slnx -c Release --no-restore -v minimal

Run the stdio host:

dotnet run --project .\src\McpServer.Host\McpServer.Host.csproj -c Release

If you want the server to operate on this repository instead of the application-owned default workspace, set an explicit root first:

$env:MCPSERVER__WORKSPACE__ROOTPATH = "D:\2026 Projects\McpServerRepo"
$env:MCPSERVER__WORKSPACE__ALLOWEDROOTS__0 = "D:\2026 Projects\McpServerRepo"

By default, Shell, Ssh, WebAccess, and Ollama are disabled in src/McpServer.Host/appsettings.json.

Client setup

LM Studio

The repository includes helper scripts for installing and validating LM Studio MCP configuration:

• scripts/install-lmstudio-mcp.ps1
• scripts/Disable-LmStudioSandbox.ps1
• scripts/Enable-LmStudioAgentMode.ps1
• scripts/Test-LmStudioAgentMode.ps1
• scripts/Test-LmStudioWorkspaceAccess.ps1

Typical install:

.\scripts\install-lmstudio-mcp.ps1 -WorkspaceRoot 'D:\2026 Projects\McpServerRepo'

VS Code / Copilot-style MCP clients

Install a user-level mcp.json entry with:

.\scripts\install-vscode-mcp.ps1

This avoids checking a machine-specific .mcp.json into the repository.

Local Ollama path

If you want local-model assistance through MCP, enable McpServer:Ollama and use the inference.local_* tools described below.

The repo includes dedicated setup/details here:

• Ollama Local Inference Tools
• Codex / VS Code MCP Setup

Local inference workflows

When McpServer:Ollama:Enabled is turned on, the server exposes:

• inference.local_status
• inference.local_complete
• inference.local_summarize
• inference.local_code_review
• inference.local_plan

The activity tools are designed to work with this model-assisted flow:

1. Call workspace.status.
2. Call activity.route with the user request.
3. Optionally call activity.context.preview.
4. Call activity.run to get a bounded context packet plus a structured-output schema.
5. Send that payload to the local model yourself.

The server prepares the model call payload; it does not invoke the model for activity.run.

Remote and web automation

When enabled in configuration, the server can expose:

• shell.exec for local non-interactive commands.
• ssh.execute and ssh.write_text for remote profile-based automation.
• web.fetch_url and web.search for allowlisted HTTP/HTTPS retrieval.

Command-style tools prefer explicit command plus args input and reject shell-control operators in raw command text.

Packaging

src/McpServer.Host is packable as a dotnet tool:

• Package ID: McpServer.Host
• Command name: mcpserver

Pack locally with:

dotnet pack .\src\McpServer.Host\McpServer.Host.csproj -c Release

Validation

The repo-level baseline is:

dotnet restore .\McpServer.slnx
dotnet build .\McpServer.slnx -c Release --no-restore -v minimal
dotnet test .\tests\McpServer.UnitTests\McpServer.UnitTests.csproj -c Release --no-build -v minimal
dotnet test .\tests\McpServer.IntegrationTests\McpServer.IntegrationTests.csproj -c Release --no-build -v minimal

There is also an inference smoke-test harness for exercising registered MCP tools through an OpenAI-compatible endpoint:

pwsh -File .\scripts\Invoke-InferenceToolSmokeTest.ps1

Documentation

• Architecture
• Method Summary
• Workspace Configuration
• Ollama Local Inference Tools
• Codex / VS Code MCP Setup
• Changelog

License

This project is licensed under the MIT License. See LICENSE.