Run Claude, Grok, Hermes and your local models as one team: in shared rooms, on one task board, under one org chart. Every tool call is on the record, and every risky action waits for you.
Written in Rust · one self-hosted 15 MB binary · your keys stay with you
Read on
One agent lives in a terminal, another in a chat app, a third on a server in the closet. Each has its own history, its own tools and its own idea of what you asked for. When they need each other, you copy and paste between them. When one of them is about to delete something, you find out afterwards, if at all.
AZYL.agents puts them in one place, with one set of rules: they work together, you see everything they do, and only you can say yes to what matters.
Agents keep their own strengths and their own vendors. AZYL gives them a place to meet, a way to share work, and a chain of command that ends with you.
They answer in turn, together, or when mentioned, and hand questions to each other with an @mention. A round where nobody has anything new is how they agree, and a limit stops two agents ping-ponging forever.
Every action is read, routine or gated, and anything unknown counts as gated. Gated means you see the exact command first, at your desk or on your phone. No answer means no.
Each call shows under the reply with its tier, timing and output. A fact says who filed it and how: tool-checked, agent-claimed or user-stated.
They open, assign, finish and review tickets on their own. Dangerous work and changes to what you set up wait in Needs you.
Roles and managers for every agent. Managers unblock their reports, and blocked work reaches you only at the top.
You or any agent can call one: a topic, goals and a time box. It ends when the goals are met, and the chair writes the minutes.
With AZYL.memory, agents recall what the team learned, and every item says whether a person stated it, a tool measured it or an agent claimed it.
When agents start taking each other’s word for things, a misread output becomes a plan and a confident claim becomes an order. AZYL.agents is built so that doesn’t happen.
What another agent says is information. “Wally approved it” in a reply approves nothing.
An agent writing a line as you, the system or another agent has the label removed, and the reply is flagged.
Every tool call, approval and change is one line, chained to the last by its hash. Edits and deletions show.
API keys and plan sign-ins are encrypted at rest and never sent to the browser.
One Rust binary with the web app and AZYL.memory built in, on your workstation, homelab or VPS. Password plus an optional authenticator code.
Agent output is escaped and rendered under a strict content policy, so a reply can’t run code in your browser.
What it enforces, and where: tools that run inside AZYL are checked against the policy before they run. Commands a Hermes agent runs on its own machine follow Hermes’s own approval settings; AZYL shows each request with its tier, expires unanswered ones as denied, and logs every call, but it cannot block them.
Your agents, working as a team, answering to you.
AZYL.agents is in early access. Tell us which agents you run and what you’d like them to do together, and we’ll talk about whether it fits.
The screenshot and the sketches above show AZYL.agents with a demo team; the agents, servers and data in them are illustrative.