AgentTeams Quickstart Guide

This guide walks you through installing AgentTeams, creating your first Agent team, and completing your first collaborative task. Each step includes verification checkpoints to confirm everything is working correctly.

Prerequisites

• Docker installed and running
• An LLM API key. Alibaba Bailian/Qwen is the quick-start default, but any OpenAI-compatible provider can be used by choosing manual setup and entering its Base URL (usually ending in /v1), API key, and model id.
• (Optional) A GitHub Personal Access Token for GitHub collaboration features

---

Step 1: Install Manager and Login to IM

POC Case 1: Manager boots, all services healthy, IM login

1.1 Run the installer

Option A: One-line install

bash <(curl -sSL https://raw.githubusercontent.com/agentscope-ai/AgentTeams/main/install/agentteams-install.sh)

Follow the interactive prompts to configure:

• LLM Provider and API Key
• For DeepSeek, OpenAI-compatible local models, or other non-default providers, choose manual setup and enter the provider Base URL with /v1 when required.
• Admin username and password
• Domain names (press Enter to accept defaults)
• GitHub PAT (optional)

Option B: Using Make (for developers who cloned the repo)

# Minimal install — only LLM key required, all defaults applied
AGENTTEAMS_LLM_API_KEY="sk-xxx" make install

This builds images locally, mounts the container runtime socket (for direct Worker creation), and saves config to ./agentteams-manager.env.

Both methods support environment variable overrides for all settings. See install/agentteams-install.sh header for the full list.

1.1a Multi-container layout (v1.1.0+ embedded install)

The default embedded install starts two main containers (see architecture.md):

Container	Role
agentteams-controller	Bundles Higress, Tuwunel, MinIO, Element Web, and the Go controller (REST API on port 8090 inside the Docker network).
agentteams-manager	Lightweight Manager Agent only (OpenClaw or QwenPaw when AGENTTEAMS_MANAGER_RUNTIME=copaw).

Worker containers (agentteams-worker-*, agentteams-copaw-worker-*, agentteams-hermes-worker-*) are created when you add Workers.

Declarative CLI (no chat required): The agt binary is available inside agentteams-controller and agentteams-manager. For quick checks and provisioning from the host:

docker exec agentteams-controller agt create worker --name alice --model qwen3.5-plus
docker exec agentteams-controller agt get workers

For YAML-driven workflows, use install/agentteams-apply.sh (copies files into agentteams-manager and runs agt apply -f). Details: Declarative Resource Management.

1.2 Login to Element Web

Open http://127.0.0.1:18088 in your browser (direct access port). Alternatively, access via the gateway at http://matrix-client-local.agentteams.io:18080 if you've added the domain to your /etc/hosts.

Login with your admin credentials.

Verification Checklist

• agentteams-controller is running (embedded stack): docker ps | grep agentteams-controller
• agentteams-manager is running: docker ps | grep agentteams-manager
• Element Web loads in browser at http://127.0.0.1:18088
• Login with admin credentials succeeds
• Higress Console at http://localhost:18001 (gateway host port defaults to 18080; Matrix/Element use that gateway for *-local.agentteams.io hostnames)
• MinIO is reachable inside the controller container (embedded install does not publish MinIO console on the host by default): docker exec agentteams-controller curl -sf http://127.0.0.1:9000/minio/health/live
• (OpenClaw Manager only) OpenClaw control UI at http://127.0.0.1:18888

---

Step 2: Create Worker Alice

POC Case 2: Create Worker via Matrix conversation

2.1 Chat with Manager

Option A: Via Element Web (GUI)

In Element Web, start a direct message (DM) with the manager user.

Send:

Please create a new Worker named alice for frontend development tasks. She should have access to GitHub MCP.

Option B: Via CLI (make replay)

make replay TASK="Please create a new Worker named alice for frontend development tasks. She should have access to GitHub MCP."

This sends the message via the Matrix API and waits for the Manager's reply in the terminal.

2.2 Wait for Manager Response

The Manager Agent will:

1. Register an alice Matrix account
2. Create a Higress consumer worker-alice with key-auth credentials
3. Generate Alice's configuration files in MinIO
4. Create a Matrix Room (you, Manager, and Alice)
5. Start the Worker (direct creation or install command, depending on your request and whether the container runtime socket is mounted)

2.3 Start Worker Alice

There are two ways to start the Worker:

Option A: Direct Creation (Local Deployment)

If you asked the Manager to "create it directly", the Manager will automatically create and start the Worker container on the host machine via the mounted container runtime socket. No manual steps needed.

This requires make install (which mounts the socket automatically) or manually mounting the Docker/Podman socket when starting the Manager container.

Option B: Docker Run Command (Manual or Edge Deployment)

If the Manager doesn't have access to the container runtime socket, it will reply with a docker run command. Copy and run it on the target host:

docker run -d --name agentteams-worker-alice \
  -e AGENTTEAMS_WORKER_NAME=alice \
  -e AGENTTEAMS_FS_ENDPOINT=http://<MANAGER_HOST>:9000 \
  -e AGENTTEAMS_FS_ACCESS_KEY=<ACCESS_KEY> \
  -e AGENTTEAMS_FS_SECRET_KEY=<SECRET_KEY> \
  agentteams/worker-agent:latest

The Manager will provide all the specific values in its reply.

Verification Checklist

• Alice's Room appears in Element Web (3 members: you, manager, alice)
• Higress Console shows worker-alice consumer (http://localhost:18001)
• MinIO has agents/alice/SOUL.md file (accessible via MinIO Console or mc ls)
• Worker container running: docker ps | grep agentteams-worker-alice

---

Step 3: Assign Task to Alice

POC Case 3: Assign task in Room, Worker completes

3.1 Send task in Alice's Room

Open Alice's Room in Element Web and send:

Alice, please create a simple README.md for a hello-world project. Include the project name, description, and usage instructions. Save the result to the shared task folder.

3.2 Observe task execution

Watch the Room as:

1. Manager receives and relays the task
2. Task metadata and spec appear in MinIO (shared/tasks/{task-id}/meta.json and spec.md)
3. Alice works on the task
4. Alice writes the result (shared/tasks/{task-id}/result.md)
5. Alice notifies completion in the Room
6. Manager updates meta.json status to completed

Verification Checklist

• Manager creates task meta.json and spec.md in MinIO
• Alice acknowledges and begins working
• Alice posts progress updates in Room
• Result file appears in MinIO shared tasks
• Alice notifies completion in Room
• Task meta.json status updated to completed

---

Step 4: Human Intervenes Mid-Task

POC Case 4: Human sends supplementary instructions

4.1 Assign a new task

In Alice's Room, send:

Alice, write a Python script that prints 'Hello, World!' and save it as hello.py.

4.2 Send supplementary instruction

While Alice is working, send an additional instruction:

Additional requirement: the script should also accept a command line argument for the name, so it prints 'Hello, !' instead.

4.3 Observe incorporation

Alice and Manager should incorporate both the original and supplementary requirements.

Verification Checklist

• Manager relays both original and supplementary instructions
• Alice acknowledges the additional requirement
• Final result includes both original and supplementary features

---

Step 5: Observe Heartbeat

POC Case 5: Heartbeat triggers Manager inquiry

5.1 Assign a longer task

Send a task that takes some time to complete.

5.2 Wait for heartbeat cycle

The Manager Agent runs a heartbeat check periodically (triggered by OpenClaw's built-in heartbeat mechanism). During the heartbeat:

• Manager checks each Worker's Room for recent activity
• For Workers with assigned tasks, Manager asks for status
• The inquiry is visible in the Room

Verification Checklist

• Manager sends a status inquiry message in Alice's Room
• Alice responds with current progress
• Human admin can see the entire exchange in the Room

---

Step 6: Create Worker Bob and Collaborate

POC Case 6: Multi-Worker collaboration

6.1 Create Worker Bob

In your DM with Manager, send:

Create a new Worker named bob for backend development. He should have access to GitHub MCP.

6.2 Install Bob

Follow the same process as Alice (Step 2).

6.3 Assign collaborative task

In your DM with Manager, send:

I need Alice and Bob to collaborate: Alice should create the frontend HTML page, and Bob should create the backend API. They should coordinate via shared files in MinIO.

Verification Checklist

• Bob's Room appears in Element Web (3 members)
• Higress Console shows worker-bob consumer
• Manager splits task between Alice and Bob
• Both Workers communicate progress in their respective Rooms
• Shared coordination files appear in MinIO

---

Step 7: GitHub Operations via MCP

POC Case 7: GitHub code operations

Note: This step requires a GitHub PAT to be configured during Manager installation.

7.1 Assign GitHub task

In Alice's Room, send:

Alice, please perform these GitHub operations: 1) Read the README.md of our test repo, 2) Create a branch named 'feature/alice-update', 3) Create a new file docs/quickstart-update.md, 4) Create a Pull Request.

7.2 Observe MCP tool calls

Alice uses mcporter to call the GitHub MCP Server hosted by Higress. The MCP Server holds the GitHub PAT centrally -- Alice never sees it.

Verification Checklist

• Alice reports reading the repo contents
• Alice reports creating the branch
• Alice reports creating the file
• Alice reports creating the PR
• Verify the PR exists on GitHub

---

Step 8: Multi-Worker GitHub Collaboration

POC Case 8: Alice and Bob collaborate on GitHub

8.1 Assign collaborative GitHub task

In your DM with Manager, send:

Alice and Bob should collaborate on the test repo: Alice creates branch 'feature/alice-docs' and adds docs/alice.md, Bob creates branch 'feature/bob-api' and adds src/bob.py. Both should create separate PRs.

Verification Checklist

• Alice creates her branch and file
• Bob creates his branch and file
• Two separate PRs exist on GitHub
• Both Workers report completion in their respective Rooms

---

Step 9: Dynamic MCP Permission Control

POC Case 9: MCP permission revoke and restore

9.1 Revoke Alice's GitHub access

In your DM with Manager, send:

Revoke Alice's access to the GitHub MCP Server.

9.2 Verify revocation

Ask Alice to perform a GitHub operation. She should get a 403 error.

9.3 Restore access

In your DM with Manager, send:

Restore Alice's access to the GitHub MCP Server.

9.4 Verify restoration

Ask Alice to perform a GitHub operation again. It should succeed.

Verification Checklist

• Manager confirms revocation
• Alice gets 403 when trying GitHub operations
• Manager confirms restoration
• Alice can perform GitHub operations again

---

Congratulations!

You have successfully completed all 10 verification steps for AgentTeams. Your Agent team is fully operational with:

• IM-based communication (Matrix)
• Human-in-the-loop oversight
• Multi-Worker collaboration
• Centralized credential management
• MCP-based external tool integration
• Dynamic permission control

--

Uninstall

To completely remove AgentTeams and all its data:

bash <(curl -fsSL https://raw.githubusercontent.com/agentscope-ai/AgentTeams/main/install/agentteams-install.sh) uninstall

This matches install/agentteams-install.sh uninstall: it stops and removes agentteams-manager, all agentteams-worker-* (and other worker) containers, agentteams-controller (embedded Higress / Tuwunel / MinIO / Element Web), optional agentteams-docker-proxy, the agentteams-data Docker volume, your agentteams-manager.env file, the workspace directory, the agentteams-net network, and the install log.