Download
中EN
GitHub Download
Six plush teammates, one for every job

AI teammates that live on your machine.

Sorts files, runs commands and remembers how you work. One key for every major model.

Download
  • Free, open-source app
  • Pay-per-use models
  • Windows / macOS / Linux
  • App UI in Simplified Chinese

02Meet the crew

Give every bot a job.Pick one — or keep the whole crew.

Six plush characters, each with a persona, a model and a workspace. Go on, poke one — it talks back.

All-rounder

Ping 天天

Its antenna pings the moment a message lands.

Good at
Default assistant, meeting notes.
Model
claude-sonnet-5-5Example model — change it any time
Salary
≈ $5.67 / mo everyday use, estimated

Lives on your computer

Reach it from your phone

Illustrative: chatting in WeChat or Feishu · swipe to switch

  • No new app to install Message it in WeChat or Feishu. Slack is coming soon.
  • Same bot, same memory What you tell it on the go, it still knows back at your desk.
  • Hand off work on the go Online while your computer is on; for 24/7, run it in Docker.

03What it does

It does the work.And it remembers.

It acts on your computer in plain sight, and keeps its memory as Markdown files on your own disk.

01Computer usemacOS Beta

Not just chat — it gets things done.

It sees the screen, opens apps, reads and writes files, runs commands — and you can watch every step.

  • Only 24 actionsScript execution, killing processes and the like are hidden from the model.

  • Writes stay in its workspaceGoing further needs your OK — or switch to read-only.

  • Auto by default, ask-first optionalUI-changing actions like clicks and typing can wait for you.

Screen content seen during computer use is sent to the model you choose.

02Long-term memory

It remembers you — and knows what not to keep.

Three layers — bot, group and team — kept separate.

  • Plain Markdown filesOn your disk, open in any editor.

  • Secrets are refusedPasswords, keys and ID numbers are detected and skipped.

  • Long chats tidied upA context steward does it with the cheaper model you pick.

Open the Downloads folder，Select 12 invoices，Drag into “Expenses”，Needs your OK：Move 12 files to ~/Documents/Expenses，Allow，expenses.xlsx · 9 KB

MEMORY.mdExample
# About you- Prefers short answers- Daily digest at 17:00## topics/project.md- Release date: confirmed- Email password: hunter2

✗ Looks like a password — skipped

And there's more.

  • Images and video Before each run it shows you the model and the prompt, and you can switch models on the spot.
  • Three permission levels Read-only, workspace or full access, set per bot. Even with full access, recursive deletes ask first.
  • Slash commands /model, /reset and more work in the app and in WeChat and Feishu chats.
  • Spending at a glance Balance, low-balance alerts and daily usage live in the app, and each bot can get its own spending cap.
  • Two network routes Pick the international route or the one optimised for mainland China; the app never switches on its own.
  • Pictures and files over WeChat Send it pictures and files; vision models can read them, and it can send files back.

04Models & pricing

One key.Every model.

Give each bot the right brain for the job — and swap any time.

One key 70 models · 13 vendors

Write a one-line tagline for our launch.

“AI teammates that live on your machine.” — want two more options?

claude-sonnet-5-5 · input $0.63 · output $3.15 / 1M tokens

Example — change it any time in the app

Vendors and text models on Modelflare：Moonshot AI、ByteDance、Zhipu AI、Google、Alibaba Cloud、xAI、DeepSeek、Meta、Anthropic、OpenAI、MiniMax、TypeSafe；kimi-k3、claude-fable-5、glm-5.3-flash、claude-opus-5-5、gemini-3.1-flash-lite、gpt-6-astra、grok-4.6、claude-haiku-5-5、gemini-3.8-flash-high、claude-sonnet-5、gpt-6.1-terra、qwen3.8-max、glm-5.2、gemini-3.5-flash-lite、gpt-5.5、grok-4.7、glm-5.3、gpt-6-luna、claude-haiku-4-5、gpt-5.6-terra、gemini-3.1-pro-preview、grok-4.5、gemini-3.5-flash、gemini-3.8-flash、claude-fable-5-1、gpt-5.6-sol、jev-1.13、gemini-3.7-flash、gpt-6-sol、claude-sonnet-5-5、deepseek-v4-flash、gemini-3.6-flash、gpt-6.1-sol、deepseek-v4-pro、codex-auto-review、gpt-5.4、jev-1.13.0、deepseek-v4.1-flash、muse-spark-1.3、claude-opus-5、claude-opus-4-8

No monthly fee. Pay only for what you use.

Prices come from modelflare.dev/pricing, fetched Nov 1, 2026 04:21 (UTC+8), version 20260821-gemini-image-resolution-v1.

  1. AnyModelBot

    $0/ forever

    Free and open source (MIT)

  2. Model usage · billed by Modelflare

    $0.4/ 1M input tokens

    e.g. gpt-5.6-sol, default route, output $2.4 per 1M tokens; official $5 / $30

    See prices for all 41 text models →
  3. What a month costs

    ≈ $3.96/ month

    Example: 50 messages/day, 3,000 input + 600 output tokens each, gpt-5.6-sol.

    Open the calculator →

Get a Modelflare key →

05Your data

Your data.Where does it actually go?

What stays on your machine and what has to leave — both drawn out.

Your deviceLocal
Leaves your device
Channels you connect
Model providers

Read the full security notes →

06Everywhere

Reachable everywhere.One bot, one memory.

Desktop, Docker and browser — plus WeChat and Feishu in your pocket, with Slack coming soon.

  • DesktopWindows · macOS · Linux

    Windows, macOS and Linux — download and go.

    All downloads and checksums

  • DockerNAS / server

    Run it on a NAS or server, always on.

  • BrowserLocal / self-hosted

    The desktop app and Docker ship a local web console.

  • Chat appsWeChat · Feishu

    Message it from the chat apps you already use — on your phone, too.

No standalone iOS/Android app yet; on your phone, chat with it in WeChat or Feishu (Slack is coming soon). The web console is local or self-hosted — not a cloud web app.

Setup guide

Run it on a NAS or server, always on.

docker run -d --name amb -p 3600:3600 -v amb-data:/data ghcr.io/modelflare-ai/anymodel-bot
All downloads and checksums →

The desktop app and Docker ship a local web console.

  1. Desktop: tray menu → Open in browser. The console listens on 127.0.0.1 and generates a random access token on every launch.
  2. Docker: open http://<host>:3600/?token=<token>; read the token with docker exec amb cat /data/ui-token.
  3. Do not expose port 3600 directly to the internet.

Channel (iLink QR login)Scan a QR code, then chat with it right in WeChat — pictures and files too.

  1. Click the bot name at the top of the chat → info panel → Channels → WeChat → Log in with QR code.
  2. Scan the QR code with WeChat and confirm. You can also run npm run weixin:login from the command line.
  3. The login is kept in your data directory and resumes after restarts; when it expires, the app asks you to scan again.

Platform limit: after each message you send, the bot can reply at most 10 times within 24 hours.

Sample QR code — not scannable

Channel (long connection + cards)A Feishu bot with streaming reply cards and approval cards.

  1. On the Feishu Open Platform, create a custom app, add the Bot capability, and grant the permissions to read direct messages, send messages as the app and write cards.
  2. In the info panel, open Channels → Feishu, enter the App ID and App Secret, and save.
  3. Under event subscriptions, choose to receive events over a long connection and add im.message.receive_v1. Publish a version, then message the bot in Feishu.

One-to-one chats only for now; group messages are ignored.

Channel (in development)We're building it: DM your bot, or @mention it in a channel.

07Good questions

Questions.Straight answers.

Got a question? Open it up.

Didn’t find it? Dig into the docs
How is it different from Grok Bot, OpenAI dots or Claude Duet?

They are the same kind of always-on personal AI agent: it works for you in the background and checks in when a decision is yours. The difference: AnyModelBot is open source (MIT) and runs on your own computer or server. It works with GPT, Claude, Gemini, Grok and more; the app is free and models are billed per use, with no single vendor's subscription. The README on GitHub has a full comparison table. Grok Bot, dots and Claude Duet are products of xAI, OpenAI and Anthropic respectively; AnyModelBot is not affiliated with them.

Is my data uploaded to your servers?

Chats, memory, files and keys are stored on your device. Model requests go through Modelflare to the provider you choose; log retention follows Modelflare's “Data retention” doc. Screen content seen during computer use is sent to the chosen model.

Does it work offline?

No. Models come from Modelflare; local models aren't supported.

Can I use it on my phone?

There's no standalone app yet; chat with it in WeChat or Feishu. Slack is coming soon.

I don't have a Modelflare account.

At first launch, click Sign in / Sign up for Modelflare and sign up in your browser. The app connects itself when you come back, with no key to copy. On Docker or a remote server, paste an API key instead. In mainland China, pick the China-optimised route under Advanced.

Is computer use safe?

Limited to 24 whitelisted actions; runs automatically by default, switchable to ask before UI-changing actions; screen content goes to the model; try it on a dedicated machine or account first.

How do I verify a download?

Each download lists a SHA-256; use Get-FileHash on Windows, shasum -a 256 on macOS/Linux. GitHub releases also include a SHA256SUMS file you can check in one go with sha256sum -c.

Is the bot still there after I close the window?

Closing the window keeps it in the tray (optional launch at login); it's offline when the computer is off or asleep; use Docker for always-on.

Do you collect telemetry?

What we can confirm: telemetry in the computer-use driver (Cua) is turned off and magpie's auto-update is disabled. See the data-flow diagram for how model requests travel. The code is MIT-licensed — read it yourself.

Bring your first bot online.

The app is free — pay only for what you use.

Download
Deploy with Docker →

MIT MIT open source · View the source on GitHub · Third-party licenses

AnyModelBot demo

Download and installGrab the installer for your system (Windows, macOS, Linux), or run it on a server with one docker command.

Animated walkthrough; the UI is illustrative.