Emissary

Production-grade Claude agents for .NET — compile-time verified, provably safe, Native AOT.

Emissary is a Claude-native agent framework for modern C#. Tool schemas and dispatchers are generated by the compiler, agent behavior is constrained by enforceable contracts, and every run is recordable, replayable, and assertable in CI.

var options = new AgentOptions
{
    SystemPrompt = "You are a support agent. Verify identity before refunding.",
    Tools = { SupportTools.VerifyIdentityTool, SupportTools.RefundPaymentTool },
};
options.Rules.Require("refund_payment", prerequisite: "verify_identity");

var agent = new ClaudeAgent(options);
var result = await agent.RunAsync("Refund order A-1001, it arrived broken.");

Start here

Guide	What it covers
Getting started	Install, first tool, first agent, streaming
Tools and schemas	`[ClaudeTool]`, parameters, structured outputs, diagnostics
Safety and contracts	Tool contracts, taint tracking, RBAC, shadow mode, compensation
Testing agents	Record/replay, behavioral assertions, canarying, LLM-judge evaluation
Production	Hosting, human-in-the-loop, sessions, resilience, caching, cost, telemetry

The API reference is generated from the source XML documentation, and benchmarks has the measured numbers behind the performance claims.

Packages

Package	What it adds
`Emissary`	Agent runtime + the `[ClaudeTool]`/`[ClaudeSchema]` source generator and analyzer
`Emissary.Testing`	Behavioral assertions, canarying, and LLM-judge evaluation
`Emissary.AspNetCore`	SSE streaming endpoints and the human-approval webhook
`Emissary.Mcp`	Expose tools or whole agents as an MCP stdio server
`Emissary.Sqlite`	Durable suspended-run and conversation persistence
`Emissary.Extensions.AI`	`IChatClient` adapter for Microsoft.Extensions.AI pipelines

Design decisions

Emissary's architecture decisions are recorded as ADRs. The ones that explain the most about how it behaves:

ADR	Decision
0001	Claude-native, not provider-agnostic
0003	100% coverage, and what may be baselined
0005	Tag-driven, lockstep versioning
0006	Compaction runs client-side, so compacted runs still replay
0007	A failed tool tells the model less than it tells the caller
0008	Tests at the SDK boundary must use SDK-produced values