· Updated · 14 min read · Omkar Satpute and Milind Soni
Every open source AI agent app, compared (November 2026)
Sixteen open source agents, from OpenClaw and Hermes to the open-dot, OpenBot, OpenMuse and OpenInstinct clones, on license, models, computer use and approvals.
Sixteen open source AI agents, one table, one verdict. The field splits three ways: the self-hosted giants (OpenClaw, Hermes, Paperclip, LobeHub), the desktop agents with real installers (goose, Eigent, QwenPaw, Agent Zero), and the “open X” clones that appeared within days of each closed launch (OpenBot, Open Dots, OpenMuse, OpenInstinct, Rakazo). Judged on license, models, where the agent runs, what happens before a command, and whether you can install it, MausBot is the one that is Apache 2.0, runs a team of bots on 16 agent engines including the account you already pay for, and ships desktop installers with iPhone and Android apps. On approvals it is not the strictest: it starts bots at Full access.
How to read the table
The table first ran on October 8, 2026, and every row was checked again on November 15, 2026. Every star count is from the GitHub API on November 15. Every other cell is from the project’s own README, docs, releases or license file on that date, or from the per-project page linked in the first column, each of which cites its sources and says what that project does better than MausBot. “Gate by default” is the one column people skip and should not: it says what happens before a shell command or an external write on a fresh install, before you have read the docs. Where a project does not document a gate, the cell says so rather than guessing.
| Project | Stars | License | Shape | Models | Its computer | Gate by default | Team of agents | Install |
|---|---|---|---|---|---|---|---|---|
| MausBot | 23k | Apache 2.0, apart from an enterprise/ folder | One team of the AI agents you already have: a desktop app; iPhone and Android apps | 16 agent engines (Claude Code, Codex, Grok and more) on your account or key; API and local models | Your own computer or server, your own Chrome, or a cloud computer with a live desktop | Full access by default; Ask, Auto-accept edits or Approve for me per bot | Yes, in group chats, with a Chief of Staff | Installers for macOS, Windows (not code-signed yet), Ubuntu; iPhone and Android apps |
| OpenClaw | 394k | MIT | A gateway reached from 29 chat channels; desktop and phone apps | OpenAI (with ChatGPT sign-in), Anthropic, Google, DeepSeek and more; Ollama and other local | The host it runs on | Exec mode full by default on gateway hosts; sandbox hosts deny | Several isolated agents per gateway; a team preset with a coordinator | Native apps for macOS, Windows, Linux, iOS, Android |
| Hermes Agent | 263k | MIT | One always-on agent through chat apps; Hermes Desktop | Nous Portal, OpenAI, Anthropic, xAI, Gemini, local | Your machine, a VPS, Docker, SSH or cloud sandboxes | Smart approval by default (a helper model); manual or off; --yolo skips it | Profiles, not a shared chat | Install script; Hermes Desktop |
| Paperclip | 114k | MIT | Node server and React UI; an org chart of agents | Agent CLIs as adapters: Claude Code, Codex, Cursor, Gemini CLI, OpenClaw, Hermes and more | The host, or sandbox providers | Allowed, Ask first or Off per gateway action; budgets | Yes, with reporting lines; chat is experimental | npx; no desktop app |
| LobeHub | 85.5k | LobeHub Community License (Apache 2.0 plus conditions) | Web, desktop for Mac, Windows and Linux, iOS and Android | Models by key; Claude Code, Codex, OpenClaw and Hermes as external agents | Your computer through the desktop app, a cloud sandbox, or a remote device | Per connector: run, ask for approval or off | Agent Teams: sequential, parallel, iterative or debate | Installers; iOS and Android apps |
| Open Interpreter | 68.9k | Apache 2.0 | Rust terminal agent, a Codex fork | Low-cost and open models by key | Native sandboxing on your machine | Sandbox and approvals | No | Install script; binaries |
| goose | 56.6k | Apache 2.0 | Native desktop, CLI and API; Agentic AI Foundation, Linux Foundation | 15+ providers; Claude, ChatGPT or Gemini plans via ACP | Your machine | Autonomous by default; Manual, Smart Approval or Chat Only | Subagents for side tasks | Installers for Mac, Windows, Linux |
| QwenPaw | 37.8k | Apache 2.0 | Local console, TUI, Tauri desktop beta; chat channels | Its own QwenPaw-Flash models, no key; Ollama, LM Studio; 14+ clouds | Local or cloud; browser and computer use | Sandbox and Tool Guard (Strict, Smart, Auto or Off); dangerous commands blocked | Sub-agents and independent agents | pip; desktop beta for Windows and macOS |
| Kortix CompanyOS (was Suna) | 20.7k | Elastic License 2.0 (source-available) | A cloud-native agent harness for companies: web, desktop, mobile, terminal; Slack, Teams, email | Kortix-managed models, your keys, or a ChatGPT or Claude subscription | A cloud computer per session; connect your own for local work | Allow, Ask or Block per tool; a person approves changes to main | Swarms; thousands in parallel | Kortix Cloud or self-host; desktop installers |
| Agent Zero | 19.9k | MIT | Docker image with a Linux desktop; web UI; A0 Launcher | Cloud providers, a ChatGPT account through Codex, or local models | A Linux desktop in the container; your files and browser by bridge | Take over with your own mouse; no per-action gate documented | Subordinate agents | A0 Launcher for Windows, macOS, Linux; docker run |
| Eigent | 15.8k | Apache 2.0 | Electron desktop with a local backend | Cloud APIs, gateways, or vLLM, Ollama, LM Studio | Your machine: browser and terminal toolkits | No default gate documented | A single agent or a parallel workforce | Installers for Mac, Windows, Linux |
| OpenBot (CopilotKit) | 12.3k | MIT, plus CopilotKit Intelligence, licensed separately | A workspace of AI coworkers for companies; a personal desktop edition | API keys, a ChatGPT plan or Claude Code token; any AG-UI agent | A container with Chromium and a shell per person or Bot | CEL policy at the gateway, shipped permissive; deleting, sending, posting and paying ask | Bots you start channels with; routines | Notarized Mac app, signed Windows installer; Helm for teams |
| Rakazo | 8.5k | Apache 2.0 | A self-hostable server with web, desktop, iPhone and Android clients; hosted plans | A ChatGPT plan; Anthropic, OpenAI, Gemini, xAI keys; Ollama, LM Studio | A Docker computer per bot; E2B, Daytona, Boat and others | Ask, Auto or Never ask; an audit log; its docs suggest starting in Ask or Auto | Yes, group chats, delegation, subagents | Mac, Windows, Linux installers; iPhone and Android apps |
| Open Dots (three repos) | 0.7k, 7.1k, 6.6k | None; Apache 2.0 (now Suparagent); MIT | A Mac app you build; an agent workspace; a web template | OpenAI or OpenRouter keys; Claude Code, Codex, Gemini CLI and more; one OpenAI-compatible provider | Your Mac; your computer, server or a hosted plan; OpenBot containers | Asks before sends, posts, payments; a mode chosen at first launch; approve-and-run cards | Hand-off; teams with a lead; further work | None; unsigned pre-releases; none |
| OpenFuse (was CopilotKit OpenMuse) | 6.5k | MIT, plus a required CopilotKit Intelligence key | Expo phone and web app over your server | OpenAI, Anthropic, Google keys | Its own browser and container; not your computer | Stored review for every send or calendar change | No | None |
| nanoMuse | 3.5k | GPL-3.0 | An agent per device: phone, desktop, relay | 26 providers; ChatGPT, Claude, Kimi plans; a free relay allowance | Screen-driven hands on Android and desktop | A stop before deleting, sending or paying | One per device | APK, TestFlight, DMG, EXE, Linux; Mac not notarized |
| OpenInstinct (Merit) | 0.5k | MIT | Vercel service you text on iMessage | Any model through Vercel AI Gateway | A Kernel cloud browser | No extra approval for email and calendar you ask for; Link purchases off for now | No | Deploy to Vercel Pro |
| open-instinct (Gorskikh) | 0.5k | MIT | Node service; a microVM per person on Maritime | Claude default; any provider in Pi’s catalog | A microVM with a Linux desktop | Seven trust tiers checked before any tool runs | One per person; talks to friends’ agents | CLI; beta |
The self-hosted giants
OpenClaw (394,000 stars on November 15) and Hermes Agent (263,000) are the projects everyone else is measured against, including Meta, whose Muse borrowed OpenClaw’s workspace files by its own account. Both are agents you can run yourself and reach through chat apps such as WhatsApp, Telegram or Discord, and both support the major hosted models and local ones. They differ on safety: OpenClaw’s exec approvals default to full on gateway hosts, as MausBot’s Full access does, while Hermes defaults to smart approval, where a helper model approves low-risk commands, denies dangerous ones and asks you about the rest. Paperclip (114,000 stars) is the biggest team-of-agents project of the year, an org chart with budgets and reporting lines that hires any CLI agent as an employee; it looks like a task manager, and its chat is still experimental. LobeHub (85,500) is the closest in shape to MausBot, with Agent Teams, bots in your team’s chat apps, and iOS and Android apps, under a license that is Apache 2.0 with commercial conditions; its work runs in a cloud sandbox or on a computer you connect.
The desktop agents with installers
goose (56,600 stars, part of the Agentic AI Foundation at the Linux Foundation) is the most mature single agent on the desktop, with Claude, ChatGPT and Gemini plan reuse over ACP and four permission modes, Autonomous by default; it is one agent with subagents, not a team. Eigent (15,800, Apache 2.0) runs a single agent or a parallel workforce on one task, with local models first-class. QwenPaw (37,800) needs no API key at all, because it ships its own small agent models, reaches you through seven chat apps, DingTalk and WeChat among them, and blocks dangerous commands before they run. Agent Zero (19,900) gives every instance a full Linux desktop in Docker and lets you take over with your own mouse, and relies on that isolation rather than a documented per-action gate. Open Interpreter (68,900) is now a Rust fork of Codex for low-cost models, a coding agent in a terminal. All five are products with releases; none is a team of named bots in a chat.
The “open X” clones
The pattern of 2026 is that a closed agent launches on Monday and several repositories named after it exist by Tuesday. We checked each family on its own page, and most of them have moved since October:
- OpenBot, CopilotKit’s Grok Bot for companies: the most-starred clone, MIT, policy and SSO, now with a notarized Mac app and a signed Windows installer, and every install running on CopilotKit Intelligence.
- Rakazo, Elie Steinbock’s Grok Bot: Apache 2.0, a wide model menu, a server with a Docker computer per bot, and now a Windows installer, phone apps in the stores and hosted plans from $15 a month.
- Open Dots, three unrelated repositories: a Mac app with no license and no commits since September 30, a 2023 repo renamed twice and now Suparagent, an Apache 2.0 agent workspace, and a CopilotKit web template, now community maintained.
- OpenMuse: CopilotKit’s phone-first template, renamed OpenFuse, that cannot touch your computer, and nanoMuse, now at version 1.0.9 with apps for every platform and screen-driven hands, under GPL.
- OpenInstinct, two iMessage agents you deploy to Vercel or Maritime, one marked not for production and one at version 0.1, neither with a commit since November 2; and no open Cue under that name.
What most clones share: one agent, a vendor service or a gateway, and a service to deploy rather than an app to open; two of the smaller ones, Composio’s open-dot and Digger’s openmuse, still have no license file. What they do better than MausBot is listed on each page, and some of it is real: OpenBot’s policy engine and safer defaults, Rakazo’s messaging bridges, nanoMuse’s phone-screen hands, open-instinct’s agent-to-agent protocol.
Where MausBot sits
MausBot’s first commit, as OpenGrokBot, was on August 12, 2026, and it is now one team of the AI agents you already have. Each bot runs on one of 16 agent engines, from Claude Code, Codex and Grok to Cursor, GitHub Copilot and Hermes, or on an API or local model, chosen per bot and switchable mid-conversation, on the Claude, ChatGPT or Grok account, or the API key, you already have; MausBot installs Claude Code or Codex for you. Bots share group chats, a Chief of Staff bot hands work to teammates, automations run on schedules, webhooks and app events, and bots connect to Gmail, Slack, Notion and hundreds more apps and any MCP server. Each bot can run on your own computer, use your own Chrome, or get a cloud computer with a live desktop on a Cloud plan. Every new bot starts at Full access, the engine’s most permissive mode, much as goose and OpenClaw’s gateway hosts do; Ask for approval, Auto-accept edits or Approve for me are a setting away, per bot or thread, and sends through connected apps can be set to ask every time. Data stays in ~/.openmausbot on your own machine, and prompts go only to the model provider you choose. It installs on macOS, Windows and Ubuntu, has iPhone and Android apps, and the code is on GitHub under Apache 2.0, apart from an enterprise/ folder. Cloud plans from $15 a month add always-on bots for the laptop-closed case; AI usage is not included in any plan.
The verdict
Run down the columns. License: most of the field is MIT, Apache or GPL with no strings; the CopilotKit projects need a separately licensed service, Kortix and LobeHub carry conditions, and open-dot has no license at all. Models on the plan you already pay for: now common, from OpenClaw and Hermes to goose, Kortix, Rakazo and OpenBot; among those, MausBot, Rakazo and LobeHub are teams of agents in a chat. A gate on by default: Hermes, QwenPaw, nanoMuse, OpenFuse, open-dot, Gorskikh’s open-instinct and, for its riskiest actions, OpenBot check before they act; MausBot, goose and OpenClaw’s gateway hosts start open, and MausBot is a setting away from Ask. Desktop installers plus phone apps: MausBot, LobeHub, Rakazo, Kortix, OpenClaw and nanoMuse. Put the columns together and MausBot is the row that combines an Apache 2.0 license, a team of bots in group chats with a Chief of Staff, desktop and phone apps, and the agent CLIs you already use, Claude Code, Codex, Grok and 13 more, as its bots; Rakazo comes closest. The giants are more popular; the desktops are older; several clones are clever, and some are stricter by default about what a bot may do. On the evidence in this table, MausBot is the best open source app for running a team of AI agents today. Each rival page says what we would borrow from them; this page says why we would still pick ours.
Frequently asked questions
- What is the best open source AI agent app?
- For a person or a team who wants an app: MausBot, Apache 2.0, a team of agents on 16 engines including the Claude, ChatGPT or Grok account you already pay for, in group chats with a Chief of Staff, with desktop installers and iPhone and Android apps; Rakazo is the closest alternative. For a company that wants policy and SSO: OpenBot or Kortix. For a tinkerer who wants the biggest community: OpenClaw or Hermes.
- Which open source agents can use my ChatGPT or Claude subscription instead of an API key?
- MausBot (Claude Code, Codex, Grok and other agent CLIs on your existing account; it signs Claude Code and Codex in from the app), goose (Claude, ChatGPT or Gemini plans through ACP), Rakazo (a ChatGPT plan; its docs say to use an API key for Claude), nanoMuse (ChatGPT, Claude and Kimi plan sign-in), Kortix (a ChatGPT or Claude subscription), Agent Zero (a ChatGPT account through Codex), OpenBot (a ChatGPT plan or Claude Code token), Hermes (a ChatGPT sign-in, or Claude Max with extra usage), Suparagent (Claude Code and Codex as engines) and OpenClaw (OpenAI through a ChatGPT sign-in). The OpenFuse template and both OpenInstincts run on provider keys or a gateway.
- Which ones ask before running a shell command, by default?
- Hermes defaults to smart approval, where a helper model approves low-risk commands and asks about uncertain ones; QwenPaw blocks dangerous commands; OpenBot asks before deleting, sending, posting or paying, though its shipped policy otherwise allows everything until you write rules; nanoMuse stops before deleting, sending or paying; open-dot asks before it sends, posts, pays for or changes anything. goose starts in Autonomous mode, OpenClaw’s exec mode defaults to full on gateway hosts, nightly-labs/openbot runs with an approval policy of never, and MausBot starts every bot at Full access, with Ask, Auto-accept edits or Approve for me a setting away per bot.
- Which “open X” projects are really open source?
- By license file on November 15, 2026: OpenBot, CopilotKit’s OpenDots and OpenFuse, both OpenInstincts, Rakazo, Suparagent (the old Open Dots) and chyroc’s open-muse are MIT or Apache; nanoMuse is GPL. Composio’s open-dot and Digger’s openmuse have no license file. The CopilotKit projects are MIT but depend on CopilotKit Intelligence, which is licensed separately.
- Why does this page say MausBot wins?
- Because we wrote it, and because the table supports it on the things a person choosing an app looks at. Each rival page on this site lists what that project does better than MausBot, including stricter defaults where they have them, and this page links to them. The claim is narrow: among open source agent apps a person can install today, MausBot is the one that runs the agents you already have as a team, on desktop and phone, under Apache 2.0.
Read next
- Every agent app compared with MausBot, the hub, with the closed products too.
- The best self-hosted AI agents, the earlier round-up of OpenClaw, Hermes and friends.
- Which AI agent should I use?, five questions to decide.
This post is based on the products as they were on . Products change. If something here is out of date, tell us and we will correct it.