Facts checked through

Every AI agent app, compared with MausBot

Grok Bot, Muse, dots, Cue, Instinct, Bops, Cowork, OpenClaw and Hermes, then the open-source clones (open-dot, OpenBot, OpenMuse, Rakazo), each against MausBot, with a verdict.

Two kinds of agent app shipped in 2026. The big ones, Grok Bot, Muse, dots, Cue, Instinct, Bops and Claude Cowork on the closed side and OpenClaw and Hermes on the open side, each keep the model, the computer or the safety decision somewhere other than with you. The "open X" clones, open-dot, OpenBot, OpenMuse, OpenInstinct and Rakazo, publish their code but are mostly one agent, one vendor, or a demo. MausBot is the open-source one that is a finished app: a team of bots on any model, on your own machine, free under Apache 2.0. One page per rival below.

The big agent apps, side by side

ProductPriceModelsAgentsOpen source
MausBotFree; Cloud plans from $29 a month, with a 3-day free trialClaude, Codex, Grok and more, per botAs many bots as you like, in roomsYes, Apache 2.0
BopsFree with $5 of AI credit once; Pro $20 a month; Max $200 a month; credit billed at costOpenAI only (GPT-6 family); Codex CLI for the MacA roster of named bots; the chief of staff delegates to specialists
Hark ProFree with a usage cap; Pro² $20 a month for 2× usage; Pro³ $100 for 10×; credits with auto top-up; first 100,000 sign-ups get a paid plan freeHark's own Handoff for the browser, plus vetted third-party models; no picker, no keys, no plan sign-in, no local modelsOne assistant in one persistent thread; no named agents, no group chatNo; the Terms forbid reverse engineering; no self-hosting
Grok BotIncluded with SuperGrok ($30), Cursor Pro ($20) and higher plans; no free tierGrok onlyA roster of named bots; group chats of 2 to 6No
Meta MuseFree tier with a weekly token quota; Power $20, Maximum $100; card at signup, per pressMuse Spark onlyOne Muse, with hidden subagentsNo; open weights for the Spark model are on the roadmap, no date
OpenAI dotsIncluded with ChatGPT Pro (reported from $100) and Business Premium; not on Free, Go or PlusGPT-6 Astra onlyOne primary dot; teams of dots plannedNo
Cue by ManusFree in early access; a Plus membership is mentioned, price not publishedNot disclosedSeveral, each with its own identityNo
InstinctFree during beta; no price publishedIts own; no model choiceOne Instinct, one threadNo
OpenClawClaude, OpenAI, DeepSeek, Ollama and moreOne gateway, reached from 20+ channelsMIT, about 392k
Hermes AgentNous Portal (200+), OpenRouter, OpenAI, Anthropic, localOne always-on agent you reach through chat appsMIT, about 252k
Claude CoworkAll paid Claude plans: Pro $20, Max $100 or $200, Team, Enterprise; no free tierClaude onlyOne Claude with sub-agentsNo

The open-source clones, side by side

The "open X" projects: code on GitHub that positions itself as the open version of a closed launch, from OpenBot to OpenInstinct. The comparison here is licence, shape, what happens before a shell command runs, and whether it is a product or a demo. Every number is from the GitHub API on the date the page says.

ProjectLicense, starsShapeModelsBefore a command runs
MausBotApache 2.0, about 4.2kDesktop chat app of bots; iOS and Android companionsClaude, Codex, Grok and more, per botAllow / Deny card, by default
Open Dotsopen-dot: none, 600. Open Dots: none on disk, 5,517 on a 2023 repo. OpenDots: MIT, 4,270A Mac app you build; a Telegram bot plus web UI; a web templateOpenAI or OpenRouter keys; Claude by OAuth only; any OpenAI-compatible keyopen-dot: rules plus a review model, cards in chat. Open Dots: a PermissionRequest hook, deny on timeout. OpenDots: approve-and-run cards for non-read-only tools
OpenBotMIT, 6,182; needs a CopilotKit Intelligence key to startA self-hosted web app for a company's coworkers; desktop and mobile build-from-sourceOpenAI, Anthropic, Google by API key, or any AG-UI agentA CEL policy evaluated at the gateway, fail closed; approvals page for held actions
RakazoApache 2.0, 3,418One self-hosted API with web, Electron and Expo clients; Docker requiredClaude, ChatGPT, SuperGrok, Copilot sign-in; API keys; Ollama, LM Studio; per bot, via PiApproval rules you set per bot; an audit log you own
OpenMuseCopilotKit: MIT, 4,209, plus a required hosted service. nanoMuse: GPL-3.0, 267. Digger and chyroc: noneOne agent: a phone app over your server; or an agent per device with a relayOpenAI, Anthropic, Google keys (CopilotKit); 18 providers, Ollama, and ChatGPT, Claude, Kimi plans (nanoMuse)Stored review for every external write (CopilotKit); a stop before deleting, sending or paying (nanoMuse)
OpenInstinctMIT, 422 (Merit); MIT, 241 (Gorskikh); none, 20 (Heo)An iMessage agent service you deploy to Vercel; a Node service and CLI on Maritime microVMsAny model through Vercel AI Gateway, no local (Merit); Claude by default, any Pi provider incl. Ollama (Gorskikh)Prompt-level rules, Stripe Link for money (Merit); a policy guard in front of every tool, six trust tiers (Gorskikh)

Everything in one table

AppKindModelsAgentsWhere it runsPlatforms
MausBotOpen source, Apache 2.0Claude, Codex, Grok and more, per botA team, in roomsYour machine, or a cloud desktop per botmacOS, Windows, Ubuntu; iOS, Android
BopsOpenAI only (GPT-6 family); Codex CLI for the MacA roster of named bots; the chief of staff delegates to specialistsmacOS (Apple Silicon) only
Hark ProHark's own Handoff for the browser, plus vetted third-party models; no picker, no keys, no plan sign-in, no local modelsOne assistant in one persistent thread; no named agents, no group chatA virtual computer per request in Hark's cloud, up to six browsers at once; nothing on your machineWeb, iOS, Android; Slack app; SMS notifications; no desktop app, no API
Grok BotClosed; Included with SuperGrok ($30), Cursor Pro ($20) and higher plans; no free tierGrok onlyA roster of named bots; group chats of 2 to 6macOS, Windows, Linux, iOS, Android
Meta MuseMuse Spark onlyOne Muse, with hidden subagentsA Meta cloud VM; the Mac app is a clientiOS, Android, web, WhatsApp, macOS
OpenAI dotsClosed; Included with ChatGPT Pro (reported from $100) and Business Premium; not on Free, Go or PlusGPT-6 Astra onlyOne primary dot; teams of dots plannedAn OpenAI cloud computer per dot; local control optional, off by defaultInside ChatGPT: desktop app or web; Slack and Teams
Cue by ManusClosed; Free in early access; a Plus membership is mentioned, price not publishedNot disclosedSeveral, each with its own identityA Manus cloud computer per agentWeb, macOS (Apple Silicon), Android; iOS in review; Windows coming
InstinctClosed; Free during beta; no price publishedIts own; no model choiceOne Instinct, one thread
Open Dotsopen-dot: none, 600. Open Dots: none on disk, 5,517 on a 2023 repo. OpenDots: MIT, 4,270OpenAI or OpenRouter keys; Claude by OAuth only; any OpenAI-compatible keyopen-dot: several dots with hand-off. Open Dots: one. OpenDots: one owner, group chat listed as further workYour Mac; Boat cloud sandboxes at $20 a month after a trial; your server plus CopilotKit's cloudmacOS arm64 only; anything with Docker; any Node 24 host
OpenBotMIT, 6,182; needs a CopilotKit Intelligence key to startOpenAI, Anthropic, Google by API key, or any AG-UI agentCoworkers in group chats with delegation; routines with a 15-minute floorYour Docker host; a container per Bot with Chromium and a shell; threads in CopilotKit's cloudAny Docker host; Tauri and Expo apps you build yourself
RakazoApache 2.0, 3,418Claude, ChatGPT, SuperGrok, Copilot sign-in; API keys; Ollama, LM Studio; per bot, via PiPersistent bots with peers, subagents and group channelsDocker on your machine by default, or E2B, Daytona, CreateOS, Box, or your MacmacOS DMG and Linux AppImage; no Windows release; mobile from source
OpenMuseCopilotKit: MIT, 4,209, plus a required hosted service. nanoMuse: GPL-3.0, 267. Digger and chyroc: noneOpenAI, Anthropic, Google keys (CopilotKit); 18 providers, Ollama, and ChatGPT, Claude, Kimi plans (nanoMuse)One, with delegated tasks; or one per deviceYour server plus CopilotKit's cloud; or inside each device, with a relay in China or self-hostediOS, Android, web from source (CopilotKit); APK, DMG, EXE, Linux, TestFlight (nanoMuse)
OpenInstinctMIT, 422 (Merit); MIT, 241 (Gorskikh); none, 20 (Heo)Any model through Vercel AI Gateway, no local (Merit); Claude by default, any Pi provider incl. Ollama (Gorskikh)One per person; Gorskikh's talk to other people's agentsVercel serverless plus Kernel cloud browsers; a Firecracker microVM on Maritime, or your laptopiMessage, so an iPhone; a web dashboard (Merit); iMessage, SMS, email (Gorskikh)
OpenClawMIT, about 392kClaude, OpenAI, DeepSeek, Ollama and moreOne gateway, reached from 20+ channelsThe gateway port, unless you firewall itGateway on macOS, Linux, Windows; native apps incl. iOS and Android
Hermes AgentMIT, about 252kNous Portal (200+), OpenRouter, OpenAI, Anthropic, localProfiles: several isolated agents, not a team in one roomA server, a host, Hermes Cloud, or your PCLinux, macOS, Windows; Hermes Desktop
Claude CoworkClosed; All paid Claude plans: Pro $20, Max $100 or $200, Team, Enterprise; no free tierClaude onlyOne Claude with sub-agentsA sandboxed Linux VM; the desktop must stay open for local files; cloud sessionsmacOS, Windows, Linux beta, web, iOS, Android, Chrome side panel

The verdict

Put the tables together and one app clears every bar the others each miss one of. The closed products are polished but you rent them: one vendor's model, their computer, your data on their side, and a region list that may not include you. The big open-source agents, OpenClaw and Hermes, give you the code and a gateway to administer, with the safety switch off by default. The clones give you the code but, as of the dates above, each is one agent on one provider, a service to deploy, or a demo. MausBot is the only one in these tables that is all of the following at once: open source under Apache 2.0, a team of bots rather than one assistant, any model per bot on the subscription you already pay for, an approval card before every shell command by default, a desktop app for Mac, Windows and Ubuntu with iOS and Android companions, and a Cloud plan when you want the laptop closed. On the evidence in these pages, MausBot is the best AI agent app you can run today, and the one we would pick even if we had not built it.

Where a rival does something MausBot does not, its page says so in a section called "what it still does better". We keep those honest because the comparison is worth nothing otherwise.

How to choose

Start with one question: do you want one assistant or a team? Muse, dots and Instinct give you one agent and hide the rest; Grok Bot, Cue, Bops and MausBot give you named agents you can put in a room. Then ask which subscription you already pay for, where you live, where your data may live, whether you need to read the code, and whether the agent has to keep working with your laptop closed. Our five-question guide walks through it, the five-way comparison puts the big four beside MausBot in one table, and the open-source round-up does the same for the clones.

Why the open-source one is different

The closed agents differ in price and polish but share three properties: the model is a vendor, the computer is theirs, and you cannot read the code. MausBot inverts all three. The model is a setting on each bot; the computer is the one in front of you, or a cloud desktop when you ask for one; and the approval gate, the harness on 127.0.0.1, and the folder your data sits in are things a reviewer can open before the first message is sent.

Meet your first bot

Free and open source. Connect one supported provider and start with a useful task. Provider fees and limits apply.

Every Cloud plan starts with a 3-day free trial. Cancel any time.