Super Agent Party: a self-hosted AI companion that also drives your desktop
⭐ All-in-one AI companion! Super Agent Party = Self hosted neuro sama + openclaw! ⭐ 全能AI伴侣！超级智能体派对 = 自托管neuro sama + openclaw!
At a glance
- What is it?
- Super Agent Party bundles a VRM/Live2D avatar, an IM and livestream bot layer, MCP tools and desktop control into one Windows-first self-hosted stack under AGPL-3.0. The scope is unusually wide, and the README answers fewer operational questions than the feature list suggests.
- Who is it for?
- Adopt Super Agent Party if you want a single self-hosted process that puts a VRM or Live2D character, an IM bot and an MCP tool surface behind one interface, and you are willing to run it on Windows 10/11 or Windows Server 2025. Do not adopt it if you need a documented Linux server deployment, a stable plugin API you can build a product on, or a permissively licensed codebase, because the project is AGPL-3.0 and the README is written for end users rather than integrators.
- Can I use it commercially?
- Yes, with strict conditions. AGPL-3.0 is a network copyleft licence: if people use a modified version over a network, for example as a hosted service, you must offer them its source code under the same licence.
- Is it still maintained?
- Yes. The repository last received commits 24 days ago.
- What is it written in?
- Mainly JavaScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What Super Agent Party is trying to replace
Most people assembling an AI companion end up wiring four separate things together: a chat front end, a TTS and avatar layer, a bot bridge for Discord or Telegram, and some agent runtime that can actually touch files or the desktop. Super Agent Party collapses those into one application. The README describes it as an "All-in-one AI companion" and as a self-hosted combination of neuro-sama and openclaw, which is a fair summary of the ambition: a character you talk to, plus an agent that does work.
The target user is not a backend team. It is someone running this on their own Windows machine who wants a desktop companion with a custom VRM model, a VTube Studio link, tavern character cards, long-term memory, and group chat with several characters at once. The same install also exposes an OpenAI-compatible API and an MCP interface, so the companion can be driven from outside. That dual identity (toy on the surface, agent runtime underneath) is the thing to evaluate.
The mechanism: one process, many surfaces
From the repository layout and README, the architecture is a local server with a browser-rendered UI plus native desktop windows. The desktop companion supports custom VRM models, actions and 3D scenes, and the project maintains a separate live2d extension repository, sap-live2d. A second path controls Live2D models inside VTube Studio, with custom actions and expression control, so the avatar can be rendered either by SAP itself or by an external application.
The agent side is layered. A Task Center runs agents in the background, described as supporting MCP and Agent Skills. Computer control combines desktop vision with a toolchain for mouse, keyboard and terminal, and the README's example is an AI playing Wordle. An embedded AI browser gives the agent its own browsing surface. An extension system loads extensions either in a separate window or in a sidebar, and the README points at a sap-extension-creator skill so an agent can generate new SAP extensions. Release notes for v0.4.1 mention ACP protocol integration with CLI agents and multi-instance extensions, which suggests the extension host is treated as a supervised process rather than in-page script.
Getting it running on Windows
The recommended path is the portable package. Download super-agent-party-win-v0.4.3.zip from the v0.4.3 release, unpack it, then double-click 一键启动(start).bat to launch. Updates come from double-clicking 一键更新(update).bat, which the README says synchronizes to the latest version of the repository, so you do not wait for a repackaged desktop build. The stated requirement is Windows 10/11, Windows Server 2025, or later.
There is also an installer, Super-Agent-Party-Setup-0.4.3.exe, with the note to choose install for the current user only. Users in China are given ModelScope mirrors for both artifacts.
If you would rather have an agent do the work, the README provides a single command:
curl -fsSL https://raw.githubusercontent.com/heshengtao/super-agent-party/main/README_AGENT.md
That file, README_AGENT.md, is the document CLI agents are told to read before guiding a human through installation. It is the closest thing to an installation specification in the material, and it is referenced separately from the main README, which is a signal that the human-facing README is not the authoritative setup guide.
Where the documentation thins out
The README is a feature gallery. It shows screenshots of the chat interface, the task center, the browser, the extension window and the multi-role group chat, and it lists integrations: QQ, WeChat, Feishu, DingTalk, Telegram, Discord and Slack for IM; Bilibili, YouTube and Twitch for livestreaming. What it does not give is a configuration reference. There is no table of config keys, no list of supported model providers, no description of where data and credentials are stored, and no statement of memory or GPU requirements for running a VRM model and vision-based desktop control at the same time.
The release notes hint at the shape of the state: v0.4.2 mentions migratable data, which implies there is a data directory you can move, and ZeroBox Security, which is named but not explained. v0.4.3 adds multi-account support and a desktop AI Dynamic Island. If you need to know exactly which files to back up or which ports are opened, the material here does not tell you. Treat the Notion English guide and the Feishu Chinese guide as required reading before deployment, and expect to read the source for anything the guides omit.
The desktop control trade-off
Computer control is the feature that separates this from a chat wrapper, and it is also the one with the clearest cost. The README says agents act "through desktop vision combined with a toolchain for controlling the mouse, keyboard, and terminal". That is a broad grant of authority to a model-driven loop, and the failure mode is not a crash. It is an agent that misreads a screenshot and clicks the wrong control. Vision-driven GUI automation is slower and less reliable than a scripted API call, and it degrades when the UI changes, when a dialog appears, or when the screen is scaled.
v0.4.1 mentions screen range control, which reads as a scoping mechanism to limit what the agent can see or touch. If you plan to leave this running unattended, that scoping is the setting worth understanding first, along with whatever ZeroBox Security does. Neither is documented in the README beyond the release note line, so verify both against your own risk tolerance rather than assuming the defaults are conservative.
Licence and the cost of keeping up
The project is AGPL-3.0. For a personal desktop companion that you run locally, the practical effect is minimal. For anyone who wants to embed this in a hosted service, or to modify it and expose the modified version over a network, the AGPL's network-copyleft terms are the thing to read before writing code. This is not legal advice; if the deployment is commercial, get a lawyer to read the licence, not a review.
Maintenance cost is driven by release cadence. Three releases appear here between May and August 2026, each adding integrations rather than consolidating: WeChat and enterprise WeChat bots, ACP protocol integration, Moss TTS, Vtube Studio control, then multi-account and a desktop dynamic island. That pace means integration surfaces move. Anything you build against the extension system or the MCP interface should be checked against the release notes on each upgrade, and the 一键更新(update).bat flow makes it easy to pull changes you have not reviewed.
How it compares with SillyTavern plus a bot bridge
The obvious alternative for the chat half is SillyTavern, which the project itself lists as a topic and whose character cards it supports. The difference in approach is that SillyTavern is a front end: it renders characters, manages prompts and context, and connects to a model backend, but it does not drive a mouse, run background tasks or ship an IM bot. To match Super Agent Party you would pair it with something like an agent runtime and a separate bot framework, and you would own the glue between them.
That is the real trade. Super Agent Party gives you one install and one update button, at the cost of a large surface you did not choose and cannot easily replace piece by piece. SillyTavern gives you a narrower tool with a clearer boundary, at the cost of assembling the companion and agent halves yourself. If your interest is the character layer, the narrower tool is easier to reason about. If your interest is an agent that can act on your desktop while wearing a face, this project is doing the integration work for you.
Editorial conclusion
Adopt Super Agent Party if you want a single self-hosted process that puts a VRM or Live2D character, an IM bot and an MCP tool surface behind one interface, and you are willing to run it on Windows 10/11 or Windows Server 2025. Do not adopt it if you need a documented Linux server deployment, a stable plugin API you can build a product on, or a permissively licensed codebase, because the project is AGPL-3.0 and the README is written for end users rather than integrators. Before committing, read README_AGENT.md rather than the main README, confirm which model providers and API keys the current release expects, and check whether the desktop control and computer vision features require permissions you are prepared to grant.
Community notes