agentd-core

agentd as a library: the agentic loop, the supervisor that owns it, durable workflows, the A2A surface, and the config system — everything the agentd binary is, minus the binary.

Most people want the binary. Reach for this crate when you need agentd's loop inside your own process — an embedded agent in a larger service, a custom front end over the same runtime, or tools registered in Rust rather than behind an MCP server.

[dependencies]
agentd-core = "2.0"

use agentd::tools;

// Register a tool the model can call, in-process — no MCP server needed.
tools::register("deploy.status", "Current deployment status", |_args| {
    Ok(serde_json::json!({"env": "prod", "healthy": true}))
});

See examples/embedded-agent.rs in the repository for a complete one.

What you get

• The ReAct loop — think, call a tool, observe, repeat, until an answer or a budget, with a context window that compacts rather than truncates.
• A supervisor holding no model. Lifecycle, limits and the process tree live in a component that cannot be prompted; the reasoning runs in child processes it can always kill. Cancellation is killpg, not a dropped future.
• Durable workflows — a graph checkpointed before every effect, so a restarted process rebuilds in-flight runs from the store and continues.
• MCP and A2A, implemented by rmcp and a2a-rs over agentd-net's credentialed transport.
• A security posture that is checked, not documented — tools tagged untrusted-input / sensitive / egress, with the lethal trifecta refused at startup; secrets by reference only, never inline, never logged.

Features

Capability is a compile-time decision. tls is on by default; a2a, workflow, cron, metrics, otel, oauth, aauth, hot-reload, config-watch, cel and exec are opt-in, and a flag whose feature is absent exits 2 loudly rather than being ignored. Building needs a C toolchain (cmake, a C++ compiler) for the protocol SDKs; MSRV is 1.96.

Full documentation: https://agentd.dev/docs/. Licensed under Apache-2.0.