<div align="center">
<h1>
<img src="docs/assets/autoswarm/AutoSwarm-logo.png" width="48" height="48" align="absmiddle" alt="AutoSwarm logo">
AutoSwarm: Unify All Your Agents. Break Context Barriers.
</h1>
[简体中文](README.zh-CN.md)
[](https://arxiv.org/abs/2610.33048) [](./COMMUNICATION.md) [](./COMMUNICATION.md)
[](https://github.com/HKUDS/AutoSwarm/releases/latest/download/AutoSwarm-Desk-Setup.exe) [](https://github.com/HKUDS/AutoSwarm/releases/latest/download/AutoSwarm-Desk-mac-arm64.dmg) [](https://github.com/HKUDS/AutoSwarm/releases/latest/download/AutoSwarm-Desk-mac-x64.dmg)
</div>
**New:** **AutoSwarm Desktop Released** — All Your Agents in One Private Workspace
One window to coordinate agents, share context, review approvals, and track progress.
Watch all your agents work together for you. [What AutoSwarm does](desktop/README.md)
<div align="center">
<a href="https://trendshift.io/repositories/82651?utm_source=repository-badge&utm_medium=badge&utm_campaign=badge-repository-82651" target="_blank" rel="noopener noreferrer"><img src="https://trendshift.io/api/badge/repositories/82651" alt="HKUDS%2FAutoSwarm | Trendshift" width="250" height="55"/></a>
</div>
<p align="center">
<img
src="docs/assets/desk/AutoSwarm_Desk.png"
alt="AutoSwarm Desk"
width="900"
/>
</p>
## Why AutoSwarm?
**No single agent can do everything. You don’t need another agent—you need all your agents working as one.**
Different agents come with different:
- 🛠️ **Capabilities** — coding, research, browsing, office work, writing, computer use, and social networking;
- 🧠 **Context and Memory** — each agent maintains its own context, with little or no sharing across agents;
- ⚙️ **Tools and Workflows** — each agent follows its own interaction patterns and execution process;
- 🌐 **Data and Application Ecosystems** — agents are connected to different platforms, services, files, and data.
Today, you are forced to switch interfaces, repeat information, coordinate subtasks, and manually combine results.
**AutoSwarm turns fragmented agents into one coordinated team.**
Give AutoSwarm a task once. It unifies all your agents into one coordinated team:
- **Selects the right agents** based on their capabilities;
- **Shares context and memory** across agents;
- **Connects complementary tools** and workflows;
- **Bridges data** and application ecosystems;
- **Runs tasks in parallel** for faster execution;
- **Coordinates agent collaboration** and keeps everyone aligned;
- **Verifies and synthesizes results** into one coherent answer.
<div align="center">
## 🤖 Agents AutoSwarm Can Call
<table width="100%">
<tr>
<td align="center" width="25%"><p><img src="docs/assets/agents/codex.svg" width="48" height="48" alt=""><br><strong>Codex</strong></p></td>
<td align="center" width="25%"><p><img src="docs/assets/agents/claude-code.svg" width="48" height="48" alt=""><br><strong>Claude Code</strong></p></td>
<td align="center" width="25%"><p><img src="docs/assets/agents/opencode.svg" width="48" height="48" alt=""><br><strong>OpenCode</strong></p></td>
<td align="center" width="25%"><p><img src="docs/assets/agents/cursor.svg" width="48" height="48" alt=""><br><strong>Cursor</strong></p></td>
</tr>
<tr>
<td align="center" width="25%"><p><img src="docs/assets/agents/copilot.svg" width="48" height="48" alt=""><br><strong>GitHub Copilot CLI</strong></p></td>
<td align="center" width="25%"><p><img src="docs/assets/agents/muse.svg" width="48" height="48" alt=""><br><strong>Meta Muse Code</strong></p></td>
<td align="center" width="25%"><p><img src="docs/assets/agents/codebuddy.svg" width="48" height="48" alt=""><br><strong>CodeBuddy</strong></p></td>
<td align="center" width="25%"><p><img src="docs/assets/agents/hermes.svg" width="48" height="48" alt=""><br><strong>Hermes</strong></p></td>
</tr>
<tr>
<td align="center" width="25%"><p><img src="docs/assets/agents/nanobot.png" width="48" height="48" alt=""><br><strong>Nanobot</strong></p></td>
<td align="center" width="25%"><p><img src="docs/assets/agents/dsh.svg" width="48" height="48" alt=""><br><strong>DeepSeek Harness</strong></p></td>
<td align="center" width="25%"><p><img src="docs/assets/agents/grok.svg" width="48" height="48" alt=""><br><strong>Grok Build</strong></p></td>
<td align="center" width="25%"><p><img src="docs/assets/agents/kimi.png" width="48" height="48" alt="Kimi Code logo"><br><strong>Kimi Code</strong></p></td>
</tr>
</table>
</div>
AutoSwarm works across Windows, macOS, and Linux. Use AutoSwarm through:
- **Desktop apps** — [AutoSwarm Desk for Windows and macOS (private assistant and live collaboration)](desktop/README.md)
- **Command-line interfaces (CLI)**
- **IDE extensions**
- **Skill-enabled agent environments**
The main host needs Skill support and local command execution. Collaborators need compatible local CLIs; your assistant can help connect missing ones.
## 🖥️ AutoSwarm Desk App
<p align="center"><img src="docs/assets/desk/desk-en.png" width="880" alt="AutoSwarm Desk app"></p>
## 🚀 Easy to use
One Lightweight Plug-in. Any Agent as Your Entry Point. All Your Agent at Work.
### How to Use AutoSwarm
Give AutoSwarm to any agent you already use and ask:
```bash
Install or update AutoSwarm for this agent. Reuse my existing configuration and verify which locally installed agents are available.
```
Your assistant will report which collaborators are ready. To add another, tell it the agent's name; it will handle setup and verification.
Then start any task with:
```bash
Use AutoSwarm to read these 100 documents and summarize the key approaches and disagreements.
```
AutoSwarm selects, connects, and coordinates the right agents for every task.
## 🤖 AutoSwarm in Action
<div align="center">
<!-- GitHub-native video attachment, hosted in https://github.com/HKUDS/AutoSwarm/issues/67. -->
https://github.com/user-attachments/assets/89a077a9-168e-441f-8aac-48de321aa968
</div>
## 📰 News
- **AutoSwarm Desk for macOS is out**: the same app on Apple silicon and Intel Macs, Python included. [Apple silicon](https://github.com/HKUDS/AutoSwarm/releases/latest/download/AutoSwarm-Desk-mac-arm64.dmg) · [Intel](https://github.com/HKUDS/AutoSwarm/releases/latest/download/AutoSwarm-Desk-mac-x64.dmg) · [Guide](desktop/README.md)
- **AutoSwarm Desk for Windows is out**: a desktop app for AutoSwarm. Talk to a private assistant, start conversations with any of the twelve supported agents, follow how they share the work, and answer their approvals in one window. One installer, Python included, automatic updates. [Download](https://github.com/HKUDS/AutoSwarm/releases/latest/download/AutoSwarm-Desk-Setup.exe) · [Guide](desktop/README.md)
- **AutoSwarm is live!**: Full technical details and multi-agent coordination algorithms are available in the [archive](autoswarm-v1.0.0/ARCHIVE.md).
## 🗺️ Roadmap
-🔌 **Expand Integrations** — Support more agent products, models, tools, and native capabilities.<br>
-🤝 **Improve Coordination** — Strengthen context sharing, communication, task routing, and parallel execution.<br>
-🛡️ **Enhance Reliability** — Improve setup, monitoring, error recovery, verification, and overall user experience.
## 🌟 Citation
If you find this work helpful to your research, please kindly consider citing our paper.
```bibtex
@misc{li2026autoswarm,
title={AutoSwarm: Automating Agent Swarms},
author={Zhonghang Li and Zongwei Li and Chao Huang},
year={2026},
eprint={2610.33048}
}
```
## For agents: How to use AutoSwarm
<details>
<summary>Installation and updates — instructions for the configuring agent (click to expand)</summary>
### Install
1. Inspect the current host and existing Python 3.10+ / CLI installations. Reuse native models, credentials, proxy settings and project context. Ask before installing a missing provider CLI; configure only the agents needed.
2. For a first installation, use `git clone` to obtain the official repository, then install the package and Skill for the **current host**. Reuse an existing checkout when available. Example for Codex:
```bash
git clone https://github.com/HKUDS/AutoSwarm.git
cd AutoSwarm
uv tool install --force .
autoswarm setup --host codex
```
Other host IDs: `claude-code`, `opencode`, `cursor`, `copilot`, `muse`, `codebuddy`, `hermes`, `nanobot`, `dsh`, `grok`, `kimi`. Without uv, use a dedicated Python 3.10+ environment and `python -m pip install .`; keep its interpreter available. Do not reuse the archived v1 environment.
3. Read the installed `SKILL.md` at the path returned by setup. It contains the launcher argv and usage instructions. Use that exact launcher when the App cannot find `autoswarm`. Run `agents` in the actual host environment; if needed, merge an existing CLI's absolute argv into `providers.<provider>.command` in the reported AutoSwarm home/config.json. Preserve unrelated settings.
4. Verify requested providers in the actual project, replacing the example provider and directory as appropriate:
```bash
autoswarm doctor --agent codex --probe --cwd /path/to/project
```
Repeat `--agent` for several providers. This spends model calls to check a reply and native resume; `verified: true` confirms the probe. Version checks only establish executable discovery. An empty native authentication list does not prove that login is required. Handle pending native approvals/questions under the installed Skill's rules, then finish verification.
Setup installs the Skill, its runtime and the host's integration. Follow the returned reload/trust instructions, then briefly name the verified collaborators and explain that the user can request others by name. If none are ready, propose connecting a suitable agent; obtain any missing installation authorization and handle configuration and verification. See [configuration and recovery](src/autoswarm/assets/references/setup.md) for native profiles, permissions and connection diagnostics.
### Update an existing installation
Obtain the requested revision from the original source. For an existing Git checkout, preserve local edits, collect active delegated work, then run on the intended branch:
```bash
git pull --ff-only
uv tool install --reinstall .
autoswarm setup --host codex
```
Use the actual host, existing AutoSwarm home and original installation destinations. Put `--home <existing-home>` before `setup` when needed; retain custom `--skill-dir`, `--settings-file` and `--plugin-dir` values. Refresh each host installation in use. For pip installations, use the existing dedicated interpreter to run `python -m pip install --force-reinstall .`, then `python -m autoswarm setup --host codex` with the same options.
Run setup **from the updated package**: the installed Skill has its own runtime copy, so updating the checkout or package alone is insufficient. The old Skill launcher cannot fetch a new version. Keep configuration, records and native authentication, reload the host as required, read the refreshed Skill and verify the needed providers. Finish with the same brief collaborator summary as installation. There is no automatic updater. See [upgrade details](src/autoswarm/assets/references/setup.md#upgrades).
### Use and recover
Follow the installed Skill to select agents, delegate ready work, collect results with `wait --any`, handle intermediate questions/approvals and verify deliverables before answering. Ordinary tasks use the current conversation; native wake-up is optional and needs a supported host binding.
Read only the relevant references: [capabilities and data scope](src/autoswarm/assets/references/agent-capabilities.md), [communication](src/autoswarm/assets/references/communication.md), [native interactions](src/autoswarm/assets/references/interactions.md), [model selection](src/autoswarm/assets/references/models.md). Diagnose retained calls and logs before retrying uncertain operations.
</details>
[Architecture](docs/architecture.md) · [Contributing and validation](CONTRIBUTING.md) · [MIT License](LICENSE)