Agentry: pricing, features, alternatives

Agentry is a Go gateway for the Agent Message Transfer Protocol (AMTP) v1.0. When an AI planner writes a workflow at runtime, Agentry coordinates execution, shares context across agents, and stores decision traces. The repo amtp-protocol/agentry was created 2025-08-22, Apache-2.0. On 2026-09-20 GitHub listed 17 stars and 5 forks. 17 stars is a heat signal, not traffic. Docs and the protocol site live at amtp-protocol.org. Requires Go 1.21+.

It sits between Kubernetes (which schedules containers, not agent graphs) and frameworks such as LangGraph or CrewAI (which keep you inside one library). The README also contrasts A2A Protocol: A2A is a wire protocol, Agentry is an execution engine plus inbox. Do not read that table as an independent audit.

Key Features

AMTP addressing: agent@domain with DNS-based discovery, federated domains, pull inboxes or push webhooks, at-least-once delivery with idempotency.
Dynamic workflows: parallel, sequential, and conditional steps from a planner. Example from the README: research across three agents, then synthesize only if confidence is above 0.8.
Context graph: traces capture what was decided, why, and which alternatives were rejected. Agents can query precedents. Handoffs keep intent and constraints. See docs/ContextGraph.md in the repo.
Context engineering: token-aware selection, summarization, and relevance filtering so the next agent does not get the whole transcript.
Ops: TLS 1.3, API keys, health checks, metrics, structured logs, Docker and Kubernetes notes in docs/DEPLOYMENT.md. Admin API is refused until an admin key file is configured.

Limitation: this is infrastructure you run. There is no hosted Agentry Cloud price on the first-party pages checked 2026-09-20. Local quick start disables TLS (AMTP_TLS_ENABLED=false), which is not a production setting. Star count is low; treat the project as early.

Use Cases

Cross-framework agents that need a shared inbox and workflow state.
Teams that want decision memory across runs, not a fresh graph every time.
Skip if you only needed A2A cards or a single LangGraph app with no federation.

Pricing

Piece	Price	Notes from first-party pages 2026-09-20
Gateway / spec	$0	Apache-2.0. About 17 stars.
Compute	Your cluster	You run the binary, Docker, or Kubernetes.
Model usage	Your provider bill	Agentry does not sell tokens.

Getting Started

git clone https://github.com/amtp-protocol/agentry.git && cd agentry && make build
openssl rand -hex 32 > admin.keys and export ADMIN_KEY=$(cat admin.keys)
AMTP_TLS_ENABLED=false AMTP_SERVER_ADDRESS=:8080 AMTP_DOMAIN=localhost ./build/agentry -admin-key-file admin.keys
Register a pull-mode agent with POST /v1/admin/agents, send /v1/messages, read /v1/inbox/user@localhost with the returned API key.

First-party start: Getting Started and the AMTP overview.

Frequently Asked Questions

Is Agentry the same as AMTP?

AMTP is the protocol. Agentry is one gateway implementation in this repo. The protocol org also points at a Python SDK from the public site.

Does it replace LangGraph?

No. The README says it is framework agnostic. LangGraph can still be an agent; Agentry coordinates messages, workflow state, and traces.

Can I run this on the public internet with TLS off?

Do not. The local example sets AMTP_TLS_ENABLED=false for a laptop demo. Production notes are in the deployment doc.

Alternatives

A2A Protocol: agent cards and tasks over HTTP, not an AMTP inbox.
LangGraph: in-process graphs, not federated agent@domain mail.
CrewAI: a Python crew runtime rather than a Go gateway.

Tips

Configure the admin key file before you expect /v1/admin to work.
Keep TLS on outside localhost.
Read docs/ContextGraph.md before you depend on precedent queries.

Conclusion

Agentry is a small Apache-2.0 gateway for people who already generate multi-agent plans and need something that will run them, share context, and remember decisions. Clone it, set the admin key, and send one local message. Do not confuse 17 stars with production proof.