10 of the Most Popular AI Tools on GitHub in 2026

We use GitHub much the same way you probably do: to find something useful, see what people are building, or try a project because it looks fun. Our writers, developers, and designers have their own interests, and over the past few days we’ve been comparing the tools getting attention right now. You might find something for your next article, video or app here, or end up following planes around a globe because you want to see how somebody built it.

Quick Answer

If you want something to explore, God’s Eye View puts aircraft, ships and public cameras on a 3D globe; for something you can make and export, VoiceStudio clones voices and dubs videos, PPT Master produces editable PowerPoints, and Diagram Design draws diagrams in your site’s style.

For coding agents, Matt Pocock’s skills and the Karpathy-inspired guidelines are good places to start when the brief or the agent’s behaviour needs tightening, while claude-mem helps carry decisions between sessions. Paperclip keeps several agents’ jobs and budgets together, ECC provides a wider development toolkit, and REA investigates how features in existing software work.

Which Tool Does What?

We’ve sorted the table by the total stars displayed on 13 October 2026, rounded to match the repository counters, so you can see the interest alongside each tool’s purpose.

There are plenty of repositories with enormous star totals, and with a newer project we’re also interested in how quickly people are finding it. REA and God’s Eye View were on GitHub Trending when we checked, which gives some context alongside the totals; the list below follows what you can do with the tools.

1. God’s Eye View

Open God’s Eye View, and you can follow an aircraft across a 3D globe, inspect its information, and bring up ships, public cameras, or weather around the area you’re looking at. But watch out for the CIA. Kidding, kidding - it’s Bilawal Sidhu’s project, and while it looks like something out of a Jason Bourne movie, there’s a surprising amount of the world you can explore through public data.

The source list helps explain what you’re seeing: OpenSky supplies flight data, AISStream supplies vessel positions, and USGS supplies earthquakes. The cameras include Tallinn’s intersections and London’s traffic cameras, among other public feeds, with some sources providing refreshed images and others supplying video. The cars moving along the roads are simulated, so that layer has a different relationship to reality from a tracked flight.

You can lock the view onto a target, follow its trail or change the display style, and the separate data layers give you something to investigate beyond how impressive the globe looks. If you’ve ever wanted to build a map with several sources of information, there’s also source code here to study and extend; if you haven’t, following planes around for an evening is still a reasonable use of it.

The Pinokio installation opens the app in your browser, and there’s a terminal route if you prefer to run the project yourself. You can start without API keys, then add providers for photorealistic 3D, additional data or voice control, with any provider costs depending on what you switch on.

2. REA

We’ve already covered REA in more detail, because understanding how an app does something is an interesting problem to hand to a coding agent. Reverse Engineer Anything connects that agent to tools for inspecting packaged applications, native binaries and behaviour, even when you don’t have the original development project.

The published Notion example gives a much better idea of this than the name alone. The agent traces a copy operation through the desktop app’s JavaScript and finds how copied blocks retain their structure: Notion keeps structured block data locally and puts a matching ID in the copied HTML, allowing a later paste to reconnect the two. That is a mechanism you can investigate and learn from, with code locations to follow when you have another question.

REA uses MCP to expose its analysis tools to agents such as Claude Code and Codex, and the terminal commands give you access to the same workflows. JavaScript and Electron packages can contain inspectable code, while compiled native programs need a route through tools such as Ghidra, Hopper or IDA to expose their instructions and reconstructed pseudocode.

With Node.js and npm installed, you can begin through:

npx rea-agents setup

Setup adds the MCP connection and workflow instructions to your chosen agent, so yeah, the appeal is being able to study a feature, an old program or a game system in enough detail to try building something of your own; what the agent can establish still depends on the files and evidence available to it.

3. VoiceStudio

If you’re making a video and need a voiceover or a dub, VoiceStudio gives you a whole lot of room to experiment without having to rely on another expensive cloud subscription. Voice cloning, voice design, transcription, dictation and audiobook production all live inside the same application, and getting started is a matter of choosing an engine, feeding it a clean reference recording, writing your script and letting the model do its work.

What makes it interesting is how much of the process you can handle in one place, especially when you're working with video and need the dialogue to land at just the right moment. The dubbing workspace gives you timed speech and transcript rows for different speakers, while its local API and MCP support open the door to connecting the whole thing to an automated production workflow, whether you're narrating an article, building voices for different characters or bringing a video to an entirely new audience.

And this is where ElevenLabs starts looking a little less essential, because VoiceStudio has the potential to replace it for much of the voiceover and dubbing work people are currently paying subscriptions for. The quality will depend on the engine you choose, of course, but the idea of having that kind of creative control on your own machine, without constantly reaching for another paid service, is pretty compelling.

The project advertises support for 646 languages and offers a catalogue of speech engines, so finding the right model comes down to what you're trying to create and which language you're working with. The open-source edition is free, while the official site offers separate commercial editions, and although NVIDIA GPUs and Apple Silicon can accelerate the process, compatible systems can also run it on a CPU at a slower pace. You can keep speech processing local without a cloud account, with optional remote services available if your workflow calls for them.

4. PPT Master

The attractive part of PPT Master is being able to hand an agent your material and receive a PowerPoint you can continue working on, with editable text and shapes, transitions, animations and support for charts and tables. Hugo He’s project is an agent workflow, so it runs through an AI tool on your machine rather than requiring you to work inside a separate presentation service.

Suppose you have a report and a deck whose design you already like: the project has routes for generating from source material, learning a reusable style, or updating an existing .pptx while preserving its design. Speaker notes can also become narration, which is useful if you want the presentation to do more than sit on a screen during a meeting.

There’s some substance behind the editable-file claim too. Its PowerPoint and SVG mapping guide explains how supported shapes and text become PowerPoint objects. It also makes a distinction worth knowing if you work with data: a chart drawn from editable shapes is different from a native chart whose data you can edit, and the native chart/table export is an explicit option.

You need Python and an agent-capable AI tool, with model usage providing the separate cost. Before giving it your own material, you can download a presentation from the example repository, open it in PowerPoint and inspect what is editable; the examples also include the SVG slides, design specifications and notes, so there’s more to look at than a finished preview.

5. Diagram Design

A diagram of how your app works can be perfectly understandable and still look out of place on your site, which is why Diagram Design caught our attention. Cathryn Lavery’s agent skill combines instructions for the diagram itself with a visual system, producing one HTML file containing SVG that you can open in a browser.

You can ask for an architecture diagram, a timeline or a sequence of events in ordinary language, and give it your site as a reference for fonts and colours. For a simple app diagram, that might mean showing how the frontend, backend, database and cache connect, with the labels and layout explaining the relationship rather than relying on a paragraph to describe every connection.

The current README lists 45 diagram types and light, dark and editorial variants. The SVG format also means the diagram consists of text and vector shapes in a file, giving you something an agent can revise when a label, connection or design choice needs changing.

The documented installation for an agent that supports skills is:

npx skills add cathrynlavery/diagram-design

You can open the resulting HTML directly in a browser, then ask the agent to revise a label or connection using the same visual style. That leaves you with a file you can keep working on as the explanation changes.

6. Paperclip

Running several agents becomes a different job from asking one to help, because you start needing to know who has a task, which result is ready to review, and where the money is going. Paperclip puts that work into an application with roles, goals, approvals, and budgets, connecting to agent software such as Claude Code, Codex, and other supported runtimes.

Its company structure is easy to picture: you can assign different responsibilities, give the agents a shared goal, and inspect their tasks and outputs from the dashboard. A developer working on implementation and an agent checking the result can have distinct jobs, while the discussion and decisions stay attached to the work they concern.

The agents can wake for assigned tasks and follow-up messages, with optional timed checks and recurring routines, so the organisation doesn’t have to mean a collection of bots constantly talking to one another. You can begin with one agent and add others when there is enough work to justify that setup.

The quickstart requires Node.js 24.11 or newer:

npx paperclipai@latest onboard --yes

Paperclip can be self-hosted, with its standard local setup using a Node.js process, embedded PostgreSQL and local file storage, while the agents still need their own runtimes and model access. Your task history and configuration live with the instance, and the model costs depend on the agents you connect.

7. Matt Pocock’s Skills

Matt Pocock’s skills are particularly relevant when an agent can write plenty of code but keeps misunderstanding what you wanted it to build. The collection breaks engineering work into practices you can use separately, including clarifying the request, producing a specification, creating tickets, debugging and reviewing the result.

One example is grill-with-docs, which interviews you about the plan and works on documentation as the decisions become clearer. Its domain-modeling instructions keep a glossary and architecture decisions, with the glossary helping distinguish terms that people might otherwise use differently. If “account” could mean a customer or a user in your app, that distinction belongs in the project before the agent builds around the wrong interpretation.

Other skills deal with the code itself, including a testing workflow and a debugging process. The value is being able to choose the part you need, then adapt its instructions to your project, rather than depending on one enormous prompt to handle every stage of the work.

The README offers plugins for supported agents and an editable-file installation:

npx skills@latest add mattpocock/skills

Include setup-matt-pocock-skills, which asks which issue tracker you use, how you label tickets and where your project documentation should live. Those choices give the later skills somewhere specific to put the work, with editable instructions you can adapt as the project develops.

8. ECC

ECC covers more of the agent’s development setup, with skills, specialist agents, commands, hooks, memory and security workflows adapted for Claude Code, Codex, Cursor, OpenCode and other environments. It is the wider option here if you want planning, implementation, review and verification to have a shared process.

The official site gives some concrete examples of what that means. Plan Canvas lets you annotate a plan in the browser and send feedback to the agent, while Unified Memory provides a way to save a handoff and search project context across supported agent environments. AgentShield examines configuration, permissions, hooks and MCP connections, which is a different job from asking the model to review the application code.

Those distinctions help explain why this is more than a bag of writing prompts. Some pieces guide the model’s reasoning, others store and retrieve project context, and others use tooling to inspect or support the environment in which the agent works.

The guided setup is documented through npx ecc-universal install --guided, with profiles and capabilities you can select for your environment. The open-source toolkit is free and separate from the paid GitHub App plans, so you can use the local workflows without subscribing to that app.

9. Karpathy-Inspired Skills

The Karpathy-inspired guidelines take a much smaller approach, with one CLAUDE.md file built around four rules for coding agents: think before coding, keep the solution simple, make focused changes and finish against a verifiable goal. Multica maintains the repository, drawing on Andrej Karpathy’s observations about the mistakes models make.

If you ask for one bug to be fixed, those rules give the agent a clear way to approach it: state the assumptions, make the change the job requires, preserve unrelated code and check whether the bug is resolved. The detailed instructions also address speculative features and unnecessary abstractions, which are familiar frustrations if you’ve watched a small request grow into a much larger rewrite.

The guidelines favour caution, with the README acknowledging that a simple typo fix shouldn’t need the full process. You can read the approach and merge useful parts with your existing instructions, giving the agent clear limits before adding more task-specific skills or tools.

10. claude-mem

Starting another session can mean losing the reasons behind a decision, even when the files themselves are still there, and claude-mem is designed to carry useful history forward. It captures observations from agent work, processes summaries and makes relevant material available when a later session needs it.

Its architecture documentation describes hooks that record tool activity, a worker that processes it and a database that stores observations and summaries. That gives the next session material about what the agent did, rather than leaving it to reconstruct everything from the current code.

The search workflow is also more selective than pasting an entire history into a prompt: it starts with a compact index, can show what happened around a result on a timeline, and then retrieves the fuller observations you need. If you want to know which fixes were tried before a feature began working, the surrounding history can be as useful as the final summary.

To begin with the documented installer, run this command:

npx claude-mem install

The project now supports several agent environments, with different installation routes and memory providers, including a hosted observer, your own OpenRouter or Gemini key, or your Anthropic plan. That provider choice determines how the summaries are processed and paid for, so a local database can still sit alongside remote AI processing.

Final Verdict

For content work, VoiceStudio, PPT Master and Diagram Design each give you something to take away and keep working on: an audio track, an editable deck or a diagram that fits your site. The developer tools need more thought about your setup, with the small Karpathy-inspired file at one end and Paperclip’s agent organisation at the other.

God’s Eye View also gives this list something to enjoy without turning it into another work task, which suits how we think about these projects. Following a plane, looking through public cameras or figuring out how the data becomes a 3D scene can be interesting enough by itself, especially if it sends you towards something you hadn’t thought of learning.