Claude Agent Csharp — AI skill for Claude Code

There's No Official Claude Agent SDK for C#. Here's What to Use Instead

Companion sample for the blog post [There's No Official Claude Agent SDK for C#. Here's What to Use Instead](https://adrianbailador.github.io/blog/71-claude-code-agent-csharp/).

Two runnable console apps, both building the same small coding agent — one that can read files, list a directory, write files, and run a shell command against a scoped workspace — using the two paths that actually exist in C# today.

Project	Approach
ToolRunnerAgent	The official Anthropic SDK's client.Beta.Messages.ToolRunner — the agent loop as an IAsyncEnumerable, one turn at a time
MeaiAgent	The same loop through Microsoft.Extensions.AI's UseFunctionInvocation(), tools defined as plain C# methods

Requirements

• .NET 10 SDK
• An Anthropic API key

export ANTHROPIC_API_KEY=sk-ant-...

Running it

dotnet run --project ToolRunnerAgent
dotnet run --project MeaiAgent

Both default to: *"Add a `.gitignore` for a .NET project to the workspace, then run `git status`"* (or *"...then show me its contents"* for the MEAI version, which doesn't wire up `run_command`). Pass your own task as arguments instead:

dotnet run --project ToolRunnerAgent -- List everything in the workspace, then read any file you find.

Each app creates a `workspace/` directory (or `$AGENT_WORKSPACE`, if set) next to the executable and scopes every file operation to it — the agent cannot read or write outside of it, `write_file` always asks for a `y/n` confirmation before it touches disk, and `run_command` (in `ToolRunnerAgent` only) checks the command's program name against an allowlist (`dotnet`, `git`, `ls`) before it even reaches the confirmation prompt.

What each project demonstrates

**`ToolRunnerAgent/Program.cs`**

• BetaRunnableTool + BetaTool — defining a tool as hand-written JSON Schema plus a Run deleg