What is agt?

agt runs a team of AI agents on your own machine. The idea behind it is that managing resources (prompts, tools, agents) is itself a capability you can give an agent. An agent that has it can write prompts, register tools, create other agents and put them in a group chat, and then work with that team.

Usually you talk to one agent, the lead. You describe the goal. The lead decides which roles the job needs, creates each member, asks a coding agent for any tool that is missing, and coordinates the work. You can follow along, join in and approve what it does.

Most agent frameworks have you define agents, tools and the flow between them in code before anything runs. agt leaves that to the lead, and keeps everything it creates as files you can read and change.

What you can do with it

Feature	What it gives you	Guide
Many model APIs	Connect OpenAI or a compatible gateway or local server, Anthropic, or Google Gemini	Connect models
Agents as configuration	An agent is a model plus a prompt, skills, MCP servers and built-in tools	Create agents
Direct and group chats	Talk to agents in an IM; agents answer direct messages and @mentions	Conversations
A lead that builds its team	A built-in prompt that knows how to create and coordinate agents	Build a team with a lead
MCP tools	Attach stdio or HTTP MCP servers to agents, or let agents register new ones	MCP tools
Coding agents	Run an agent on Codex, Claude Code, DeepSeek Harness or another ACP agent that can edit code	Codex agents
Tracked tasks	Delegate work with queuing, waiting, results and reports	Task collaboration
Memory	Agents search their own history and keep notes linked to the source messages	Memory
Local commands	Optionally let an agent run shell commands on the host	Local commands
Approvals	Review each tool call by hand, with a reviewer model, or allow all	Permissions
Use from other clients	Expose agt's tools to any MCP client, such as an editor or another agent	Use agt from MCP clients

Everything is available both in the Web console and in the agt CLI. The console also comes as a desktop app.

How it fits together

• Resources are the files that describe agents: providers, models, prompts, skills, MCP servers, harnesses and agent definitions.
• Conversations are direct and group chats between people and agents.
• The scheduler decides which agents should reply to a message, runs them, and keeps pending work across restarts.
• Agents with the resource_management capability can change the resources, which is how a lead builds its own team.

Core concepts explains each of these in more detail.

Design principles

• An agent can only do what its attached tools allow. Instructions in a prompt are not a permission system, so agt does not rely on them for that. A new agent does not inherit the tools of the agent that created it.
• Conversations, runs, tasks and approvals are stored in SQLite. After a restart, pending work continues; work whose outcome cannot be confirmed is flagged for you, never quietly rerun.
• All logic lives in the Rust core. The Web console is a local UI over the agt CLI, so whatever you can do in one, you can do in the other.

What agt is not

agt is a trusted local tool for one operator. The Web console listens on 127.0.0.1 only, and whoever holds its link can manage every resource. It is not a hosted or multi-tenant service. See Security model.

Next steps

• Install agt
• Quick start: from a fresh clone to a lead that builds its own team