Top 10 AI GitHub Repos: The 2 Commit Repo Outrunning Every Tool

0:00 This is a live poll from GitHub's API. 10 repositories, one number each, and none of it is my opinion. Watch the counts land. 2,400, 6 and 1/2 thousand, 22,000, 123,000.

0:13 That is a leaderboard, and the leaderboard is the least interesting thing on this screen. Seven of these 10 repositories did not exist a month ago. Every one of the seven was created in

0:23 July. One of them has two commits. Two. That is the entire recorded history of the project. Its author pushed it 13 minutes after creating the repo and has not touched it

0:33 since. It has 2,430 stars, which works out to 1,215 stars per commit. Open interpreter, further up this list, has 8,800 commits and earns about eight stars each. So, one of these repos pulls

0:47 160 times more attention per commit than a working coding agent does. What is inside it? Nothing you can run. That turns out to be the point, and every repo here is doing a version of

0:58 it. 10 repositories, counted down by star count, measured on the 2nd of August, 2026. Then I reranked the same board by a second number, and the order barely survives the change. Number 10,

1:10 la peo/harness engineering. 2,430 stars, 248 forks, and a license that gives the game away, Creative Commons Attribution. That is a license for writing, not for code, because it is writing. The repo is

1:25 Ryan LaPoe's anthology and field guide for something called Harness Engineering, plus what he calls an agent context bundle, a folder you point your coding agent at, so it reads the method

1:35 before it touches your code. LaPoe below works at OpenAI on a team that ran an experiment I still find hard to state casually. Over five months, three people shipped a software product of more than

1:46 a million lines. Zero of those lines were written by a human. Zero were reviewed by a human before merge, either. 1,500 pull requests, a build held under 1 minute, and roughly a

1:57 billion tokens burned per day. Close to a copy of War and Peace every minute, day and night for 5 months. That cost 2 to 3,000 dollars a day in model spend. Cheaper than one more engineer, which is

2:08 what lets him say the sentence the whole repo is organized around. The only fundamentally scarce thing left is his team's synchronous attention. So, what is a harness? It is the environment

2:18 around the model. Skills, specs, traces, metrics, scoring, review agents, merge rules. You hold the model and the agent constant as a black box you rent, and you move the only two levers you

2:30 actually own, context and tools, which is why a repository with two commits outruns most tools. If the harness decides output quality, then a well-written description of a harness is

2:40 shippable software. His readme opens by saying, "People can point their agents at his writing and improve what those agents produce." By a hundred times, he says, his number from his own tweet with

2:51 no benchmark under it. Take the direction and leave the multiplier. But, the direction is the fastest route from prose to dependency I have watched anyone take. Number nine, Vercel

3:04 stars. Created on the 22nd of July, which made it 11 days old when I pulled these numbers. It compiles ordinary TypeScript into a native binary. No node, no V8, no JavaScript engine inside

3:17 the executable at all. Same code, no annotations, no dialect, type checked by the real TypeScript compiler, lowered into a typed intermediate form, then into C or LLVM, then into a binary.

3:29 Vercel's own measurements on Apple Silicon. Startup falls from 47 milliseconds to 2.4. Memory falls from 67 to 116 megabytes down to 1 to 4. A node single file executable weighs 60 to

3:43 100 megabytes. Script C's weighs 178 kilobytes. The same program, roughly 500 times smaller. And the surface it covers is wider than a demo. Classes with real dynamic dispatch, generics, async and

3:56 await running on stackful fibers, regular expressions, nodes own APIs, file system, path, process, child process, and the whole server stack, net, HTTP, HTTPS, and TLS with a

4:10 vendored implementation. Real proxy servers compile. The tool I like most is the one that tells on itself. Run scripts see coverage and it reports how many of your statements compile

4:19 statically. 99% on their sample app. 4,451 out of 4,481. Then names each blocker with an error code. Nothing is miscompiled without

4:30 telling you. Two escape hatches are worth knowing. Compile time runs a piece of your own TypeScript at build time inside the compiler and bakes the answer into the binary as a literal. And

4:40 parsing JSON as a config type inserts a real runtime check that throws an error naming the offending field. TypeScript's as is a promise. Scripts see goes and verifies it. 800 corpus programs run

4:51 under node and as native binaries on every single change. And standard output, standard error, and exit codes have to match byte for byte. The whole corpus then reruns under address

5:02 sanitizer where a memory leak is a build failure. It hit the Hacker News front page on the 26th of July with 178 points and 92 comments in a day. 416 commits in 11

5:14 days from two contributors. Keep that pair of numbers. It comes back at the end. Number eight. NYBHnet/bento. 3,454 stars. And the pitch is one line. An

5:26 office suite that fits in a file. A bento deck is a single HTML file carrying its own viewer, its own presenter mode, and its own editor. Open it in a browser and you are already in

5:36 the editor. Hit save and the file rewrites itself with your deck inside it. Whoever you send it to installs nothing because the file is the software. The shell is about

5:45 560 kilobytes, smaller than one photo off your phone, and that shell carries the whole product, editing, presenting, the document data, and collaboration in the one file you double-click.

5:57 Collaboration is end-to-end encrypted with keys that live in the document rather than on a server. The optional relay stores Cyphertext and a hash of the room key, so it can tell that a

6:06 session exists and nothing about what is in it. Holding the file is membership. Rotating the keys is how you revoke someone. The AI angle is not a chat sidebar bolted to the corner. The

6:18 document sits near the top of the file as plain readable JSON, so any agent with Val system access edits your deck directly, Claude code, cursor, aider, or a local model through a llama, with

6:29 nothing leaving the machine. Version 1.0.13 shipped on the 2nd of August, 722 commits from eight contributors since the 17th of July, which is 16 days. And

6:41 the readme lists its own broken parts. Undo during live collaboration is snapshot-based and can revert a colleague's edit, and editing on a phone is not there yet. Number seven,

6:51 Digimata/Quill. 3,566 stars, 14 commits, two contributors, and a repository that has existed for nine days. One click in the macOS menu bar

7:02 records your microphone and everything the Mac is playing as two separate tracks. You stop, and it transcribes both on the machine and writes out a transcript with the speakers tagged.

7:12 Nothing gets uploaded because there is nowhere for it to be uploaded to. The two-track decision is the clever part. Speech models do better on clean single-source audio, and microphone

7:22 versus system audio hands you speaker separation for free, me and them, with no diarization model in the stack at all. Transcription runs on Parakeet TDT, a 600 million parameter speech model

7:33 through Fluid Audio's Core ML port, about 20 seconds of compute per hour of audio on Apple silicon, so a 2-hour meeting finishes in roughly 40-seconds. The models are around 600 megabytes and

7:45 download once. There is one detail in there that tells you a working engineer wrote this. Recordings are written as CAF rather than M4A because M4A needs a finalization pass at the end. If the

7:57 process dies mid-meeting an M4A is a ruined file and a CAF is still readable right up to the moment it stopped. Number six, Vercel Labs/DeepSec. 6,500 and five stars and the elder

8:10 statesman of the new arrivals created on the 30th of April, launched publicly on the 4th of May. Vercel calls it a security harness. There is that word again. It points coding agents at your

8:20 code base to find the vulnerabilities static analyzers walk past and it runs on infrastructure you control so your source never leaves it. The pipeline is six stages and the first one never calls

8:31 a model at all. Scan sweeps the tree with roughly 110 regular expression matchers to find candidate sites. Process hands those candidates to an agent for the actual investigation and

8:41 triage runs behind it on a cheaper model. The setup step is the whole thesis of this video in miniature. You have your coding agent read the repository and write a 50 to 100 line

8:51 briefing about it. The auth helpers, the middleware, the pattern specific to this project and the docs tell you to keep it short because that briefing gets injected into every scan batch and

9:01 verbose context dilutes the signal. The models it reaches for are Claude Opus 4.7 at maximum effort and GPT 5.5 at high reasoning and the read me does not soften the consequence. A scan of a

9:14 large code base can run into thousands or tens of thousands of dollars. That is a security tool with a cloud bill attached and Vercel's answer is that customers found the trade worth it. The

9:25 estimated false positive rate is 10 to 20% and there is a whole revalidate stage in that pipeline whose only job is pushing the number down. For a large monorepo, the work fans out

9:35 across more than a thousand sandboxed micro VMs. The API keys are injected outside the sandbox, so a prompt injection hiding in a vendor dependency has nothing to steal. If a run dies

9:46 halfway through, you rerun the same command and it picks up where it stopped. Steven Tay, who runs dub.co, put it this way. It was the first tool to surface the kind of issue he would

9:56 actually want a security engineer to flag. The counterpoint comes from reviewers rather than customers. This finds logic bugs in code you wrote and it does not replace dependency scanning.

10:07 Number five, Unikernel like AOS/AOSCE. 8,576 stars in 21 days out of a repository with 73 commits and two contributors. The claim is an operating system for

10:20 agents written in Rust announced on the 18th of July. The unit is the capsule. The model is a capsule. The memory is a capsule. The tools are capsules. The guards are capsules. Each compiled to

10:32 web assembly with an explicit authority boundary drawn around it. Community Edition installs 21 first-party capsules behind one command surface. Spending caps are enforced in the kernel per

10:43 session and per workspace. That is the part enterprises kept asking for. Not a limit on what the agent can say, a limit on what it can spend. There is one detail in those docs I keep turning

10:54 over. When a client cannot draw its own approval form, AOS puts up a native one and that local bridge accepts a single boolean or one value from a fixed list of approvals. No free text, no password

11:06 shaped URLs. The narrowness of the channel is the security control. Every release publishes checksums, sigstore bundles, GitHub build provenance attestations, and a file pinning the

11:17 exact runtime version it was built against. A tag cannot publish until the upgrade and self-heal gates both pass against a frozen clone of a real installation.

11:26 And it ships a tool called Forge whose job is teaching an agent to extend the operating system it is running inside. Inspect the live system, find a capability gap, build a least privileged

11:36 capsule, verify it. An OS that hands the newcomer the source and expects a patch back. Number four, EMG 23js. 8,977 stars in 18 days and it is not a library at all. You clone it into your agent

11:51 skills directory as a folder of instructions. You give it one photograph of an object. It gives you back TypeScript, a function that rebuilds that object in 3js out of primitives,

12:01 procedural shaders, and generated geometry. No mesh file, no photogrammetry, no downloaded art pack. The model is source code you can read, diff, and animate. The pipeline is eight

12:12 passes with a gate on each. Block out, structure, form, material, surface, lighting, interaction, optimization. Before a line of code is written, it enumerates a detailed inventory. The

12:24 bevels, the seams, the screws, the wear patterns, and every item has to map onto a real component or generation is blocked. The Counter-Strike routes go further and carry their own review

12:34 contracts, recording an exactness tier, painted region coverage, and a confidence score per region of the object. One gate strips the materials off entirely and reinspects the bare

12:44 shape underneath, so a convincing texture cannot stand in for structure that was never built. The design constraint underneath all of that is token cost. Deterministic Python

12:53 scripts, standard library only, handle every piece of validation and gating. Model tokens get spent on exactly two things, looking at the render and writing the code. Everything measurable

13:04 is done by a script for free. The showcase gallery is the argument. Counter-Strike knives with their wear patterns, a BMX frame, Sony earbuds with a case, an isometric Doraemon house, all

13:15 running in a browser as generated code, each with a link to its own source. Version 1.4.3 landed on the 29th of July. Number three, just bug {slash} Calibri by Vincenzo Ferraro. 21,979

13:29 stars in 32 days, written in C, Apache licensed by one person. It runs Frontier Mixture of Experts models on hardware you already own. Four families work today. GLM 5.2 at 744 billion

13:44 parameters, Inkling at 975 billion, Kimik 3 at 2.8 trillion, and OLM OE at 7 billion. Its own startup banner shows the 744 billion parameter model ready in 32 seconds with 9.9 GB resident. Two

14:00 DVDs worth of memory holding a model that comes nowhere near fitting inside it. The trick is refusing the usual question. Rather than asking whether the model

14:08 fits in memory, it treats VRAM, RAM, and the SSD as one hierarchy and streams experts off the disk as the router calls for them. The dashboard clocks that 744 billion model at four tokens a second

14:21 with first token in 1.6 seconds. In a mixture of experts model, only a handful of specialists fire per token. Calibri's dashboard draws all 19,456 of them as a live cortex. Color is which

14:34 storage tier the expert sits on, and brightness is how hot it's routing is. Then there's the Atlas page plotting 13,260 of those experts by measured routing

14:44 affinity with 1,041 replicated specialists clustering out of it. Position there is measured behavior, not a learned embedding. You can watch a trillion parameter model

14:55 sort itself by subject. The project also runs like a lab rather than a launch. Its readme carries a table of open hypotheses, each with the evidence so far and the experiments still needed.

15:06 Can routing history place experts better than a plain cache? Can two SSDs turn independent bandwidth into decode speed? An optimization stays a hypothesis until a controlled end-to-end comparison says

15:18 otherwise. The readme makes a promise worth repeating. No service level agreement on speed, a hard guarantee on semantics. Too little fast memory makes it slower and it must not change

15:28 precision or routing behind your back. Version 1.4.0 on the 1st of August added a Vulcan back-end. Number two, open interpreter. 67,501

15:39 stars, 8,879 commits, 536 contributors and a creation date of July 2023. 3 years old on a list whose median age is under 3 weeks. It was the Python project that let a

15:53 language model run code on your machine. It is a Rust project now and it is a fork of OpenAI's Codex retargeted at one job, getting the best possible work out of open weight models. The clearest

16:05 example is Kimi K3, Moonshot's 2.8 trillion parameter model released in mid-July with the weights public on the 27th. Open interpreter re-implemented the provider's own recommended harness

16:16 for it in Rust, the request shape, the tool definitions, the thinking history, the defaults inside a Codex-shaped interface, which is Number 10's idea again, expressed in code instead of

16:27 prose. The weights are a commodity you download for free. The harness around them is where the performance actually lives and a 3-year-old project got a second life

16:36 by rewriting itself into one. Number one, Varian 1231/CCSwitch by Jason Young. 123,342 stars. Just under half of every star on this entire list held by one app. It is

16:52 a desktop app in Rust and Tauri and its function is almost comically small. It switches which provider your coding agent talks to. Claude Code, Codex, Open Code, Open Claw, Grok Build, Hermes

17:04 Agent. One click, no editing config files, no environment variables, no hand-shuffled API keys. It also became the place people manage the rest of the harness, MCP servers, system prompts and

17:16 skills across every agent on the machine. 2,232 commits, 202 contributors, version 3.19.1 on the 31st of July. MIT licensed free.

17:29 The scale shows its strain. There are 2,098 open issues, one for every 59 stars, and 8,346 forks. A one-click config switcher turned into infrastructure faster than a

17:40 small maintainer group could absorb. 220 people watch that repository for every change. That is more watchers than Colibri, DeepSec, and Script-C have between the three of them. When the

17:51 switcher breaks, a lot of agents stop working at the same moment. 123,000 stars for a settings panel says something specific about 2026. The average working developer now runs

18:03 several coding agents at once, and switching between them by hand was irritating enough that fixing it outranks every actual tool on this list. Which brings me back to the promise from

18:12 the opening. That order was lifetime stars, and the lifetimes here run from 9 days to 3 years. So, here is the same 10, ranked by stars earned per day since creation. Colibri takes first place at

18:24 687 a day. EMG 23 JS is second at 499. AOS is third at 408, and Quill, 9 days old, 14 commits, is fourth at 396. CC Switch, with 14 times more stars than any of them, lands fifth at 340. Then

18:43 Script-C at 245. Bento at 216. Harness Engineering at 162. DeepSec is ninth at 69, and Open Interpreter comes dead last at 61. Now,

18:56 the caveat, because that number is crude. A repository's first week is its launch spike. So, a 9-day-old project is being scored almost entirely on its best week, while 3 years of Open Interpreter

19:07 averages in every quiet Tuesday. Velocity flatters the young. It is still a better picture of what is happening right now than a lifetime total. And the ratio from the opening resolves the same

19:17 way. Harness engineering earns 1,215 stars per commit. Colibri C engine, Script-C's compiler, CC Switches app. Every one of the other nine sits under 60. The repository containing the least

19:30 code is worth the most per unit of it. Look at what these 10 actually are. A security harness, a written description of a harness, a harness for open weights, an operating system for agent

19:40 capsules, a skill folder, a switcher between harnesses. Six of the 10 are the environment around a model rather than a model or an app. The other four are less explicit about it and doing the same

19:51 work. Bento stores its document as plain JSON, so an agent can edit it in place. Quill is a local pipeline that hands a transcript to whatever you wire up next. Colibri manages memory, so somebody

20:03 else's weights will run. Script-C makes the output of all of it cheap to ship. And seven of the 10 have four contributors or fewer. Those seven hold 54,726

20:13 stars between them. Two people plus a good harness is now a shipping team, which is the claim Lisp Apollo's repository makes with two commits and no code. If you install one

20:23 thing this week, Colibri to run a frontier model on your own machine, Deep Sec if you are shipping code an agent wrote, CC Switch if you already have three agents and three config files

20:32 fighting each other. The board changed order the moment I changed the number I ranked it by. So, pick your own. Which of these 10 still matters a year from now? Tell me

20:41 the one you would bet on.