The project's README from github.com/HKUDS/AutoSwarm. As one markdown file: README.md.
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
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:
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:
|
|
|
|
|
|
|
|
|
|
|
|
AutoSwarm works across Windows, macOS, and Linux. Use AutoSwarm through:
The main host needs Skill support and local command execution. Collaborators need compatible local CLIs; your assistant can help connect missing ones.

One Lightweight Plug-in. Any Agent as Your Entry Point. All Your Agent at Work.
Give AutoSwarm to any agent you already use and ask:
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:
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.
-🔌 Expand Integrations — Support more agent products, models, tools, and native capabilities.
-🤝 Improve Coordination — Strengthen context sharing, communication, task routing, and parallel execution.
-🛡️ Enhance Reliability — Improve setup, monitoring, error recovery, verification, and overall user experience.
If you find this work helpful to your research, please kindly consider citing our paper.
@misc{li2026autoswarm,
title={AutoSwarm: Automating Agent Swarms},
author={Zhonghang Li and Zongwei Li and Chao Huang},
year={2026},
eprint={2610.33048}
}
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.
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:
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.
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.
Verify requested providers in the actual project, replacing the example provider and directory as appropriate:
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 for native profiles, permissions and connection diagnostics.
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:
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.
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, communication, native interactions, model selection. Diagnose retained calls and logs before retrying uncertain operations.