If this was useful, we write about stuff like this every Friday in TalkHead. One analysis, one number, one company's experiment, and something from our own research.
If this was useful, we write about stuff like this every Friday in TalkHead. One analysis, one number, one company's experiment, and something from our own research.
21 AI GitHub repositories to watch in November 2026
A public PDF can become a tidy Markdown file with its table scrambled. MarkItDown converts files for text analysis; Docling also parses document structure. Either needs a check against the source before another tool relies on its output.
The 21 projects below solve different jobs, from local chat and document parsing to workflow automation and design. Each has a proposed first task, setup requirements and costs. Choose the result to verify before choosing the repository.
Download the installer for macOS or Windows from the links in the README. Linux has an install script and manual instructions.
Choose a local model that fits your machine, download it and open a chat.
First task, summarise a short public article and check the result against the original.
Needs A Mac, Windows PC or Linux machine with enough disk space and memory for the model you pick.
Costs Free to run locally. The README also mentions cloud models but does not state their pricing.
The catch Local models need disk space and enough memory. Ollama also offers cloud models, so check which mode you selected before assuming a prompt stays on your device.
Use the official website linked from the README to get the Windows, Mac or Linux app.
Connect a supported provider, or a local model through Ollama or LM Studio.
First task, give two models the same public document and compare what they miss.
Needs Windows, Mac or Linux, plus an account with a model provider or a local model.
Costs The Community Edition is free, while cloud providers charge for usage. The README lists a paid Enterprise Edition without prices.
The catch The desktop client is separate from the models it uses. A cloud provider may charge for usage and receive your prompts; a local looking interface does not make every connection local. The Community Edition uses AGPL-3.0, which the README says permits commercial use if you comply with it.
Ask a technical colleague to follow the README. It offers a pip route (Python 3.11 recommended) and a Docker route.
Connect Ollama or your chosen model provider, then open the local browser address shown in the guide.
First task, run a short question through the intended model and verify which provider answered.
Needs Python 3.11 or Docker, and a model to connect to.
Costs The project is free. The README also promotes a paid Enterprise Plan without prices. Model providers charge separately.
The catch Docker runs the application in a packaged environment; someone still owns updates, stored data and access. Most of the code sits under the project's licence, which includes a branding clause. Review it before deployment.
For the easiest introduction, use the cloud option linked from the project. For self hosting, the README gives a Docker command and opens the editor at localhost:5678.
Build a manual workflow using sample data before connecting a live account.
First task, turn three sample enquiry records into a tidy spreadsheet.
Needs A cloud account, or Docker if you host it yourself.
Costs The README does not state prices for the cloud option; check n8n's pricing page. Connected AI and other services can add charges.
The catch Self hosting brings server upkeep, credentials and backups. n8n is fair code under its Sustainable Use License, rather than a permissive open source licence.
Try Dify Cloud through the README, or have a colleague follow the Docker Compose self hosting instructions and open localhost/install.
Connect a model provider and build one simple workflow.
First task, make a draft FAQ assistant from a small set of public company information.
Needs A cloud account, or Docker with Compose v2.24.0 or later and at least 2 CPU cores and 4 GiB of RAM to self host.
Costs The README says the cloud sandbox plan includes 200 free GPT-4 calls. Model usage and hosting can cost money.
The catch Self hosting needs infrastructure and maintenance; the project's minimum server requirements are not a production capacity estimate. The licence is Apache 2.0 with added conditions, so read the LICENSE file before commercial use.
The Flowise team announced it was winding down the project. Feature development stopped on 29 July 2026, the repository became read only on 13 August, and the team’s presence on GitHub and Discord ended on 31 August. The npm packages and Docker images are marked deprecated.
The announcement names no hosted successor and encourages teams to fork the code, which stays public under Apache 2.0.
The catch For a work trial, n8n or Dify are safer starting points unless someone on your side can maintain a fork.
Check that you have Node.js 24.11 or newer, then follow the README's quick start, which runs one npx command to set it up.
Connect one supported agent before building a larger team.
First task, assign a narrow research task, inspect its output and compare the recorded work with the brief.
Needs Node.js 24.11 or newer. A local database is created automatically for testing; production use needs your Postgres.
Costs The software is free and needs no Paperclip account. The agents you connect bill their model providers, and Paperclip has budget warnings and hard stops.
The catch Paperclip coordinates agents you supply. Their models, subscriptions and infrastructure remain separate costs. Start with limited permissions and a task you can review.
Have a developer follow the README's recommended Docker run command and set the model connection.
Connect a small test application through the documented client or API.
First task, save a made up preference, start a fresh interaction and check whether the app retrieves it correctly.
Needs Docker and an API key for a language model. The README says it supports many providers, including local ones.
Costs The software is free, though the connected model can generate usage charges.
The catch This is a component to integrate, not a standalone office assistant. Stored memories need retention decisions, and the connected model can generate usage charges.
Use a Python virtual environment and install the optional dependencies for the formats you need; the README provides a pip route.
Run the command-line converter on a public file and save the Markdown output.
First task: compare headings, links and a table against the original.
Needs A supported Python installation and the dependencies for your file formats.
Costs Local conversion uses your computer. Optional Azure and model integrations may add usage charges.
The catch The output is aimed at text-analysis pipelines, not faithful page reproduction. Check tables and scanned content before trusting downstream answers.
Have a technical colleague follow the README's prerequisites and Docker guide. The README tells you to check out a specific version tag, so check it against the latest release.
Configure the required models, then import a small set of public documents.
First task, ask five questions whose answers you already know and inspect the cited passages.
Needs At least 4 CPU cores, 16 GB RAM and 50 GB of free disk, plus Docker 24 or later and Docker Compose 2.26.1 or later.
Costs The software is free. Servers and model usage cost money, and the project also offers a hosted service.
The catch Those figures are a starting point; workload and optional models can require more. Citations help you check an answer, but do not prove it is correct.
Open Releases through the README and choose the desktop package for your system, a .dmg for Mac, an .exe for Windows, or an AppImage or .deb for Linux.
Follow the platform guide and download the engine or model needed for your task.
First task, generate a short passage using a supplied voice before attempting a larger project.
Needs A desktop computer. An NVIDIA GPU or Apple Silicon speeds it up. Without a GPU it runs on the CPU, more slowly, and the setup needs about 5 GB of free disk.
Costs Free to run under the AGPL-3.0 licence.
The catch Speed and available engines depend on hardware. The README notes that Intel Macs need a remote backend. Voice and language support varies by engine, so test your exact use case.
Ask a technical colleague to follow the quick start, clone the code, download a model, build the command line tool and run it on a sample file.
Convert your audio first. The basic command line tool reads 16 bit WAV files, and the README shows an ffmpeg command to convert other formats.
First task, transcribe a short recording, then check names, numbers and quotations against the audio.
Needs Git, CMake and a C++ build setup, plus ffmpeg for converting audio.
Costs Free, and it runs on your machine. Larger models need more memory.
The catch This is an engine and set of examples rather than a polished editing studio. Larger models need more resources, and every transcript still needs checking.
Have a developer follow the CLI quick start, one command creates a project, one previews it in the browser and one renders the MP4.
Create an example project, preview it and render an MP4.
First task, change the text in a short template and check the rendered video.
Needs Node.js 22 or later and FFmpeg.
Costs The README states no price for the tool.
The catch This is a code based production route. It suits teams with a developer or coding agent more than someone expecting a conventional video editing timeline.
Use the README's installation section for the agent app you already use. It covers Claude Code, Claude Desktop, Codex, Cursor, Grok Build and Gemini, and you can also ask your agent to install it from the repo.
The tool runs through uv, and the first start downloads its CAD runtime.
First task, request a simple box with explicit dimensions, then inspect and measure the result.
Needs A compatible agent app and the uv tool, plus an internet connection on first start.
Costs The tool is free, though your agent's usage costs apply.
The catch The project sends usage statistics and crash reports by default, tagged with a random install ID. The README says file names, contents and prompts are never included, and gives an opt out command, uvx cadgen telemetry off. Generated geometry needs checking before fabrication.
Download an installer from the project's Releases page. The README lists macOS, Windows, Linux, FreeBSD and a web build.
Open a copy of a non critical image, never your only original.
First task, open a copy of a PSD, save it again and check the result in Photoshop.
Needs A desktop computer.
Costs Free.
The catch The README says it is early alpha and not yet a Photoshop replacement for daily professional work, and the project's roadmap says Photoshop does not yet reliably accept the PSDs it writes. Keep your originals.
Download the installer for your system from the latest release. The README lists Windows, macOS, Linux, FreeBSD and a browser build.
Open a short clip you have the rights to and make a simple edit.
First task, export a 30 second sequence and check playback on another device.
Needs A desktop computer. Windows and Linux get less testing than macOS, according to the README.
Costs Free.
The catch The README says it cannot open Premiere project files and has no plugin hosting. Its scorecard puts readiness for real work at about 56 percent.
Read the hardware requirements before starting. The README lists at least 8 GB RAM, with 16 GB or more better, and 22 GB free disk for the smallest model.
On Windows, download the ZIP and double click START-HERE.bat. On Mac and Linux, install the listed tools, clone the repo and run ./start-here.sh.
First task, choose the recommended model, let setup finish and try a short prompt in the browser dashboard.
Needs At least 8 GB RAM and 22 GB of free disk. A GPU is optional.
Costs Free, with model downloads using your disk and bandwidth.
The catch Disk streaming changes the hardware trade off; it does not guarantee fast answers on every computer. Each model keeps its authors' licence, and some are for non commercial use only.
Ask a developer to install it with npm. The README gives one global install command, and Git 2.41 or later is required.
Configure a model provider, or use the documented delegation mode so a supported coding agent runs the review with its model.
First task, review a small change whose intended behaviour and known defects your team understands.
Needs Git 2.41 or later, npm, and a model provider unless you use delegation mode.
Costs The tool is free, though the configured model can incur charges.
The catch The configured model can receive source code. Treat findings as review suggestions, with a developer checking both missed defects and false alarms.
Running an app yourself does not settle who can read the information it processes. Our reporting on AI mental-health services and conversation privacy examines the gap between a private-feeling interface, access to its records and responsibility for what happens next.
For the PDF trial, keep the source beside the converted text and check the table, reading order and missing content. Apply the same principle to another project’s first task: count the work and cost left after the result passes review.
Sources and selection
Original 18 entries: stars and releases checked 18 October 2026. Browser Use, MarkItDown and Docling: documentation checked 24 October; document tools rechecked 25 October. This is a curated watchlist, not a live trending ranking or a hands-on benchmark.
Check the current README and licence before installing. For general setup, costs and data questions, see the GitHub installation guide.