source-code

agentd

Run AI agents the way you run services. agentd is one small Rust binary that runs an agent from a configuration file: a one-shot job that exits with a meaningful code, or a daemon that wakes on schedules, webhooks, MCP resource changes and upstream events. Its tools come from MCP servers you name, people and other agents reach it over A2A, and the reasoning runs in child processes a supervisor can budget and kill.

Documentation · Quickstart · Why agentd · Releases

Who it is for

Teams that want agents in production with the discipline they already apply to services: reviewed configuration, bounded spend, a known blast radius, durable work and an audit trail. It fits event-driven back-office work, scheduled jobs, served workers other systems call, multi-agent teams, a coding agent you host yourself, and edge devices where the agent is the only thing the machine runs.

It is not a library you write an agent in, and not a hosted service.

What sets it apart

• No ambient tools. Every capability comes from an MCP server you declared, so a run's blast radius is the servers in its configuration.
• Reasoning runs where it can be killed. A supervisor that never talks to a model holds the state and the credentials, enforces budgets and deadlines, and kills a loop that misbehaves.
• Durable by default. Workflows and conversations are checkpointed to a store; a restart resumes work instead of repeating it.
• The right model for each job. Chat, reasoning and decision models, each bound to a role — the agent's turns, deep thinking, typed decisions, triage, approvals — with fallback chains across providers.
• Operable. Configuration is validated before anything runs, the exit code is an API, and every event is a JSON line on stderr.
• Open protocols. MCP for tools and context, A2A for people and agents, OpenAI-compatible, Anthropic and Bedrock model wires, OTLP and Prometheus for telemetry — and, as previews, the Personal Agent Protocol draft for acting at companies for a person and x402 for paying within a spend policy.

Install

curl -fsSL https://agentd.dev/install.sh | sh

Linux on x86-64 and arm64. The container image is ghcr.io/agentd-dev/agentd, and the terminal and web clients are npm install -g @agentd-dev/cli. Install covers every option, including building from source.

Quickstart

export OPENAI_API_KEY="paste-your-key-here"
agentd \
  --instruction "Explain in three sentences what the Model Context Protocol is for." \
  --intelligence https://api.openai.com/v1/chat/completions \
  --model gpt-5.1 \
  --intelligence-token '{{secret:OPENAI_API_KEY}}'

The answer prints on stdout, telemetry streams to stderr as JSON lines, and the exit code says how the run ended. The Quickstart continues from here: a tool from a public MCP server, the same agent as a daemon, and the terminal UI attached to it.

Where to go next

You want to…	Start with
decide whether agentd fits	Why agentd, Use cases, Security at a glance
build your first agents	Quickstart, Your first workflow, Build a coding agent
connect models and tools	Connect model providers, Connect MCP servers
deploy and operate it	Deploy, Monitor in production, Harden a deployment
understand how it works	Architecture, Security model
look something up	Configuration reference, CLI and exit codes, Workflow node reference
build on it	Embed agentd in Rust, Build a client

The documentation lives in docs/ and is published at agentd.dev/docs. Design decisions are recorded in rfcs/.

Crates

Crate	Responsible for
agentd-cli	The agentd binary
agentd-core	The engine: supervisor, agent loop, workflows, intelligence, A2A, state
agentd-mcp	The MCP client over the official Rust SDK, and a raw HTTP listener
agentd-net	Blocking HTTP/1.1 and SSE, unix and TLS connects, the SSRF guard
agentd-instruction	The Instruction Specification: parse, validate, deliver, verify

License

AGPL-3.0-only — see LICENSE. agentd-instruction is MIT OR Apache-2.0, and the conformance corpus under Apache-2.0; see NOTICE. Commercial licensing (for proprietary embedding or AGPL-free use) and commercial support are available: agent@agentd.dev.