If you run a small business and you already use ChatGPT or Claude, you've probably wondered what else is out there. GitHub is where a lot of new AI tools show up first, usually before anyone has made them pretty. I'm not a developer, so I read these pages the way you'd read a menu in a language you only half speak: slowly, looking for the dishes I recognise.
This is the second list of its kind. If you missed the first one, it's here: 5 AI Tools Worth Watching on GitHub This Week (No Coding Needed). All five tools below are new to this series.
A few ground rules before we start. Everything here comes from each project's own README, so it's what the authors say, not something we have tested. Star counts were copied on 10 November 2026 and they move every day. These are young projects, so expect rough edges. And 'worth watching' means exactly that: watching, not installing on your work laptop tonight.
Hex is the one I'd explain to my family first. Its README calls it “local voice dictation for Mac, Windows and Linux” and sums up the idea as “Hold a key, speak, and it pastes.” The README also says the first run downloads a speech model and that after that “nothing leaves your machine”.
Why it matters to you: if you write a lot of emails, proposals or WhatsApp replies, typing is often the slow part. Dictation that happens on your own computer also means your client's name doesn't travel to yet another cloud service.
The pull: about 6.2k stars and 374 forks, under an MIT licence. That's a lot of attention for a dictation tool. One caution from me, not from the README: dictation apps normally need microphone access and permission to paste into other apps, so read what your system asks for before you click Allow.
Poster describes itself as “Self-hosted social media automation” that turns news on your topic into posts for X and Facebook, approved in Telegram. Its README says it collects stories from RSS feeds, Google News and X profiles, an AI model picks one and writes the posts, and the drafts land in Telegram. Only after you tap approve does the bot publish.
That approval step is what I like about it. In its news workflow, the README says a draft is only published after someone approves it in Telegram, and it mentions a check that warns you when a draft looks like a recent post.
The pull is modest: 24 stars, 3 forks, under the AGPL-3.0 licence. Setup is not light either. The setup the README describes uses Docker, n8n, a Telegram bot and an AI model, and publishing through an API needs the matching developer app. The README also notes that new X developer accounts pay per usage. Think of it as a weekend project for someone technical, or a job you hand over.
37Studio's README calls it “An AI video studio and social media desk that runs on your own computer.” It says it builds a brand library from your website or a folder, makes branded videos, publishes to nine platforms, and drafts replies to messages and comments from a knowledge base. The README offers manual, approval-based and automatic reply modes, so check which one is selected before you connect an inbox.
The pull: 155 stars and 22 forks, MIT licence. It needs a Claude plan or an Anthropic API key. There is a desktop app for Windows and Linux, and the project's Docker guide asks for about 16 GB of RAM and about 20 GB of free disk.
Now the part you should read twice. The project's Docker guide says the launcher sets the agent's permissions to “bypass”, which lets the agent run any command inside its container, and that the default safe list “is not a security boundary”. It also says the agent can read the AI keys in its environment, and that the backup archive contains tokens and keys. We have not independently verified any performance claims. Interesting, yes. A tool to point at your company accounts without a technical person beside you, no.
Lampline's README says “Test and run phone agents on your own machine.” Read that carefully, because it is a testing tool first. It imports agents from Dialogflow CX, writes conversation tests, and grades test calls with an AI caller. It connects to phone lines through Twilio, Telnyx or SIP, and has a free local mode that downloads around 3 GB of models, with replies the README calls “slower and simpler”.
The pull: 457 stars, 48 forks, Apache-2.0 licence, version 0.3.0. That version number tells you how early it is, and the README lists features such as call history as roadmap items rather than shipped ones.
Two cautions. Test audio goes to speech services that may be in the cloud unless you choose local models. And the README says to call only agents you own or are allowed to test. If real customers' voices could ever be involved, talk to a professional about data protection before you go near it. We wrote about a different phone-agent project last month: 4 hidden AI tools on GitHub this week.
Scout's README says it is “an open-source agent that uses a real browser to finish a task for you.” It reads pages, clicks, types and scrolls, then gives you an answer. It also says Scout “runs best on Outcrawl's cloud”, and that the GitHub repository is an automatic mirror of Outcrawl's own.
The pull: 744 stars, 15 forks, Apache-2.0 licence. You need Node 22 or later and a model endpoint of your own, because Scout has no built-in model. The cloud browser needs an Outcrawl API key, and the README says it includes 1,000 credits a month.
This is the one I'd be most careful with. The README says runs are recorded by default with a replay link, but it doesn't say what it stores, for how long, or how it treats passwords typed during a task. Until you've read the provider's own documentation, don't point it at anything that needs a login or a sensitive form. It's also worth asking whether the websites you'd use allow automated access.
I ask three questions. Where does my data go? Who approves before something is published or sent? And how old is this thing? Stars are attention, not quality, and a project with 24 stars can be perfectly good while one with 6,200 can still ask for permissions you don't want to give.
If an answer isn't in the README, I treat it as unknown, not as fine. That's the whole method, and it works for tools far less exciting than these.
Do I need to code to use any of these?
Not to read about them, but most need Docker or a command line to install. Hex is the friendliest according to its README, which lists installers for Mac, Windows and Linux.
Are these tools safe?
We haven't tested them, so we can't say. The READMEs give some warnings, and the sensible default is to try anything new on a spare machine with no real client data.
What does self-hosted mean?
It means the software runs on your own computer or server instead of the vendor's. You keep more control, and you also take on the setup, updates and security.
Can I use AI to post or answer calls for my business?
Often yes, but rules on data protection, advertising and recorded calls apply. Ask a gestor (an administrative agent) or a lawyer about your own case before you automate anything customer-facing.
Why do you list star counts?
Because they show how much attention a project has right now. They say nothing about whether it works for you, which is why we copy them with the date.
Not sure whether any of this fits your business, or how to try it safely? Write to us on WhatsApp or through the contact form on our website.
Prefer Spanish? Read this post in Spanish: 5 herramientas de IA en GitHub que merece la pena vigilar esta semana (para pymes).
Sources: the GitHub pages and READMEs of anomalyco/hex, imdangro/poster-app, Shaitrish/37studio, C4rohan/Lampline, outcrawlai/scout, read on 10 November 2026.
