AnyModelBot puts a small team of AI bots on your own machine, keys included

The pitch on the repo page is one line: an AI bot team that runs on your own computer, any model, pay per use. AnyModelBot appeared on 27 October 2026 and had 4,929 stars by 21 November. It is a desktop application, not a service, and most of its design decisions follow from that.
The first thing you notice is plural. You do not get one assistant. You get a roster. Each bot has its own persona, a default model, a workspace folder, and a memory. They can be different models from different providers, because the app does not care where inference happens.
What a bot can touch
A bot can read and write files, run commands, and operate the desktop once you authorise it. Writes outside its workspace stop and ask you first. That one line, "it will ask before writing outside the workspace", is the load-bearing safety feature, because the alternative is an agent with your home directory and a plausible reason.
Around that sits the usual scaffolding: background tasks, scheduled rules, and a queue of items waiting for your confirmation. The README frames these as built in rather than plugins, which is the right call for a tool that expects to be left running.

Five chat apps, one roster
The bots are reachable from the app itself and from WeChat, Feishu, Telegram, Slack and Discord.
This is the feature that decides whether you use it. A desktop agent you have to open is a tool. An agent that answers in the group chat where the work already happens is a colleague. The trade-off is equally obvious: five surfaces means five places where a message can arrive, and five sets of permissions to think about. Teams that bridge a local agent into a company chat should decide early which channel is the official one.
Where the data goes
The README makes three promises: conversations, memory, files and API keys stay on your device; models run through Modelflare with pay-per-use billing after login, or through your own official API keys directly with the provider; and the app itself charges no subscription.
That combination is unusual. Most local-first agents either route everything through the vendor's billing or make you wire up keys by hand for every provider. Here the two paths coexist, and the app is explicit that it takes no cut of the second one.
The honest caveat is the one that applies to every local-first tool: the machine is the server. It needs to be on for scheduling to work, its disk holds the memory, and its network is the security boundary. If your laptop sleeps, your bots stop. If your laptop is compromised, so are the keys.
Model choice as a cost lever
Because bots carry their own default model, the roster becomes a budget. A cheap model for the bot that sorts files and drafts replies, a frontier model for the one that reviews a contract, a local model for anything that should not leave the building.
Treating the roster as a budget works better than picking one model for everything. It also creates a small discipline problem: every bot is another bill, another key, and another set of behaviours to learn. Six bots is a team. Sixteen is a hobby.
The setup follows that logic. Download the app for your platform, sign in to Modelflare or paste official keys, then create a bot by giving it a name, a persona, a workspace folder and a default model. The workspace is the part to think about before the persona: it defines what the bot can touch without asking, and a folder per bot keeps two agents from editing the same file at the same time.
Scheduled rules are the second thing to set up, because a bot that only runs when you open the app is a chat window with extra steps. A daily summary at 8am, a folder watch that files incoming invoices, a reminder that checks a page: those are the tasks that make the roster earn its disk space, and they all depend on the machine staying awake.
The 0.1.44 release tells you what people do with it
The changelog entry for 20 November fixes two very specific things: images dragged into a conversation were rejected when the extension did not match the real format (phone photos, .jfif files), and previewing a bot-generated web page could not read other files in the same folder.
Both bugs come from the same place. People are using this to look at real files and real pages, not to chat about abstractions. A bot that builds a small HTML page and cannot load the stylesheet next to it is useless for the task people actually gave it.
How it compares
The README compares itself to Grok Bot, OpenAI dots and Claude Duet, which is the right comparison set: persistent personal agents that keep working between conversations. The difference it claims is locality and model choice.
Worth being precise about what locality buys. It buys data residency, offline tolerance for local models, and no per-seat pricing. It does not buy privacy from the model provider, because the prompt still travels to whichever API you configured. If a task must not leave the machine, the bot has to run a local model, and then you are trading capability for that guarantee.
Setting expectations
Three limits to plan around.
This is young software. Version 0.1.44 means the shape is still moving, and a small team shipping weekly will break things that a bigger vendor would not.
Memory is only as good as its retrieval. Every one of these products claims persistent memory; what differs is whether the bot surfaces the right memory at the right moment. Test that with a specific question about something you told it three weeks ago, not with the demo.
Approvals are a UX contract, not a sandbox. "It asks before writing outside the workspace" is a rule the application enforces on its own actions. It is not an OS-level jail, and a determined prompt injection inside a document the bot reads is still a live risk. Keep the workspace scoped and the keys limited.
Who this is for
Someone who already pays for two or three models, has a folder of half-automated tasks, and wants one place where a few specialised bots can pick them up, with the history and the files staying put.
That is a narrower audience than "everyone who wants an AI assistant", and it is a more honest one. The product is a desk with several trays, not a magic employee. If the trays are the shape of your work, the fit is good.
Related articles
Six Amateurs, 1,113 Episodes, a Thousand Yuan of Compute an Episode: How an AI Drama Got Hand-Built
The first ten seasons of Armed Into Song run to roughly a thousand episodes and have passed ten billion cumulative views, with six core creators and not one of them film-school trained. Compute runs 1,000–2,000 yuan an episode; a normal shot takes three rolls, a battle twenty-plus, and anything that looks flat is graded a sixty and thrown out.
One server, one AI streamer: inside the code that runs a nightly livestream
open-yoki is the production code behind an AI host who streams nightly on Binance Square, with the persona, voice and money features stripped out. Headless Chrome draws the frame, ffmpeg pushes RTMP, a folder of Markdown decides what she feels and what she posts.
Pinky lets an agent drive your phone, and holds the Send button until you click Allow
Pinky connects any MCP agent to a real iPhone or Android phone and enforces approval in its own gateway rather than in a prompt. Banking apps are invisible, password fields are off limits, and on Android 13 the agent gets a hidden second screen.
ByteDance's SeedRealtime takes two frames a second and decides when to speak
Doubao-Seed-Realtime-1.0 folds camera frames, screen frames and speech into one full-duplex session, at no more than two frames per second. The interesting decisions are the frame rate, the proactive speaking, and the audit reply that ships switched on.