Show HN: Ante, a coding agent in a single binary that runs offline | Hacker News Reader

Show HN: Ante, a coding agent in a single binary that runs offline

github.com

169 pointsubermon93 comments

Linking to a github repo for a binary release (no source code related to the agent that I could see) is a bit iffy IMO. You should clarify your intentions or link to something else. Might confuse folks.

Linking to a binary is iffy from a security perspective. Linking to a GitHub repository is exactly what HN should do.

A GitHub repo with no source code is exactly equivalent to linking the binary directly

there are so much content and valuable stuff in the github repo you can pretty much recreated with your own agent.

I will leave this here: https://usehax.dev/ GitHub repo: https://github.com/OleksandrChekhovskyi/hax

This is a coding agent implementation I am working on, which delivers what this promises (at least on the "lean" part), except it's actually fully open source, and even more lean (few MBs of runtime memory usage).

MIT-licensed, written in C, multi-provider / multi-model, minimalist approach to system prompt and tools (think kinda like pi, but with a bit more "batteries included", like subagents and background tasks out of the box), polished presentation, inspectable (usable transcript view), etc.

nice, good to see more contributor in this space

we put it in the repo README, will add migrate more into public repo as soon as possible.

It has to be 100% open source public code or we will have no way of proving this is not malware, or secretly swapped out for malware later when your CI/CD system or laptop is compromised. Supply chain attacks happen all the time and with closed code no one will be equipped to spot it when it happens.

Also, security aside, engineers want the freedom to modify and experiment with the tools we rely on.

Tools like this are too important to be closed. Do you want to be Internet Explorer or Firefox?

Oh no, definitely not Internet Explorer. Will figure out a way to ship with source first to address the security and concern and then figure out how to do the open source development in agentic era later, I was too carried away by the complexity of the latter and ignorant to the former.

When it is 100% freely licensed open source software I can modify and compile myself, I will be happy to give it a shot.

Open source didn't stop Jia Tan... and most people on Earth still use a closed source, proprietary browser without thinking twice.

>Open source didn't stop Jia Tan

How do you think the backdoor situation would have been resolved if xz hadn't been open-source?

I was kind of excited for this until the binary blob. You want me to give your agent binary god access to my computer, and I am not even permitted to see the source code or use my own supply chain security hardened rust compiler stack? What a joke. Hard pass.

if you're interested in an open source agent that uses a minimal amount of CPU and RAM and has source that is easy to audit:

https://github.com/smol-env/smol

here are traces from an agentic task around using duckduckdb

comparing CPU and RAM usage of the whole container over time w/ OpenCode, hermes, pi, codex, smol

https://x.com/__tosh/status/2086882367126286466

https://x.com/__tosh/status/2086882204060160350

smol is very minimal only using stdlib (in this case it is the go version but you can also take a look at implementations in python, clojure, php)

also no need to trust these bench runs, you can just run your own

(any OpenAI Responses API compatible endpoint works, if your endpoint does not support 'custom' tools you can have your agent change the smol implementation to use 'function' tool implementation instead)

that said: be aware that smol does not come with any system prompt and does not load agents.md files by default

some older not so strong models benefit from a system prompt and guidance in agents.md that complements them

that said 2: system prompt and or loading agents.md automatically is easy to add though if you want it

> “while taking the time to figure out how open source should work in the agentic era”

I can’t even guess what this means

generally the challenge now is that 1. how to deal with PR spams by AI bots 2. how to make the project sustainable especially when one has no distribution. when anyone can insta remix and re-package and re-sell your hard work. For knowledge sharing open source it is ok, but if you are serious about what you built, this is question needs to be answered first before make it a true community effort.

there are still many bad players in the industry, i wouldn't mind sharing it with trusted group. But I am not yet strong enough with deal and handle all those yet.

I have bad news for you. No matter what you do here it is probably not sustainable. Competing commercially with people who have so much more resources and ability to build is not sustainable. If it's open source you need adoption, and shouldn't care if people want to steal and remix your work as that will drive more adoption.

But from a fellow engineer's perspective we don't need another commercial solution in this space. Every major tech vendor is working on harnesses, and we will get ones that will run circles around yours for free. The open source ones that get popular will have mit/freebsd/etc permissive licenses. (or no one will adopt them) But coming at it from the angle you are indicates you haven't reasoned well about this. Please don't hurt yourself and those around you by starting a startup on this idea.

I understand that claude-code takes a lot of memory and that's bad. However, harneses are simple loops, in theory should take very little memory even if written in python or typescript. See for e.g. pi agent

I love pi and share many vision and value with it. But my view on harness is that it is to capturing the structural mechanism with llm interacting the world. They are a dynamic duo evolving together. The technical depth will continue to grow (e.g. /goal being the new primitive, multi-agent collaboration is basic need)

Yes the core part is a simple loop, but we have all built toy compilers, inference engine, browsers (it is just a curl command eth) etc. The core algorithm is supposed to be simple.

Had a discussion recently: https://x.com/NoCommas/status/2086568454434537710?s=20

no source code

> We care about the harness, not the model or the prompts.

I wonder if this is a viable approach; after all frontier model providers are betting on the opposite.

Then again, they bundle their harness and offer subsidiary pricing - so maybe they themselves aren’t sure if models are as important.

my view on harness is that it is to capturing the structural mechanism with llm interacting the world. They are a dynamic duo evolving together. The technical depth will continue to grow (e.g. /goal being the new primitive, multi-agent collaboration is basic need) So it is here to stay. And it is just our focus as we don't have enough resource (yet) to improve the model and I think prompts belong to the user.

Had a discussion recently: https://x.com/NoCommas/status/2086568454434537710?s=20

How good is it to work on building games, compared to existing agents? I am building my own game?

we have a detailed launch thread explaining and show case exactly this! https://x.com/NoCommas/status/2086835536598351955

You aren't building your own game if you have a chatbot do it for you.

That's like saying you aren't building a house if you use a hammer to drive nails instead of your hand. AI is just a tool like any other, and you use it to build things like you would any other tool.

> That's like saying you aren't building a house if you use a hammer to drive nails instead of your hand. AI is just a tool like any other, and you use it to build things like you would any other tool.

A coding agent is more like a carpenter, a mason, an electrician,... rather than a hammer in that case.

I'd say it's closer to a Roomba then a hammer -- if you just let a Roomba run around your living room, can you say you vacuumed?

Well, you tell your Roomba "clean the floor" but you don't ask an AI "make a game". You give it very specific instructions.

Specific compared to writing procedural code> Barely even by analogy, IMO

One could even argue what defines AI instructability is heuristics as opposed to specifics

I wouldn't even compare it to a tool. People frequently "build their own house" where they sub out 75% of the skilled labor and act as a General Contractor/glorified gopher.

And you won't get purity tests from the layperson: in the end, you're responsible for the build quality so if you tirelessly labor/oversee those teams you're considered capable; if it ends sub-par, then you're a stooge.

A far more accurate analogy would be people saying someone didn't write a book if they hired a ghostwriter, which is generally acknowledged to be true. LLMs are much closer to being a ghostwriter than they are to an inert, non-powered hand tool.

> AI is just a tool like any other

Really? Last time I checked, other tools can operate without ripping off the works of millions of humans.

Same idea that sends people down needless game engine development / procrastination rabbit holes.

This is such a horrible cringe and bad faith take.

A classic tale from the music production world comes to mind:

"I thought using loops was cheating, so I programmed my own using samples. I then thought using samples was cheating, so I recorded real drums. I then thought that programming it was cheating, so I learned to play drums for real. I then thought using bought drums was cheating, so I learned to make my own. I then thought using premade skins was cheating, so I killed a goat and skinned it. I then thought that that was cheating too, so I grew my own goat from a baby goat. I also think that is cheating, but I’m not sure where to go from here. I haven’t made any music lately, what with the goat farming and all."

Does prompting an AI music generation service count as making your own songs? I noticed you left that one out of your parable.

If figma draws you a gradient, are you really a designer?

Probably yes; If Figma is deciding for you where the gradient goes, what colors it is, and what its purpose in the design is, then no.

it kinda is by showing you tools in a certain order and what defaults and so on

A designer, yes. A painter, no.

Does snapping a photo with a camera count as making your own painting?

We don’t call the end result a painting though, we call it a photo. Theres probably some value to distinguishing between them.

A set of pleasing qualia.

As a songwriter, drum hitter, and acoustic strummer, I am not good at drums (can play on beat, though), guitar (can strum all the cowboy chords, though), or singing (no caveat on this one).

So sometimes when I want to hear what my song COULD be if I were good at every one of those, I will use Suno. But I do not let it change my lyrics, or chose it's own arrangement. I will give it a VERY ROUGH demo (out of key singing, basic drums and strums, and it will "polish" my turds into really shiny turds.

No, it's using FL Studio's random sequence generator with a general MIDI instrument, exporting it, then saying it's a ballad you just composed. Or paying a ghostwriter to write on a topic, then saying it's a novel you just wrote.

"Good artists copy, great artists steal." - Picasso

> "Good artists copy, great artists steal." - Picasso

The problem being that the users of LLMs aren't even putting forth the effort to do the stealing themselves, which would at least have implied some level of discernment and artistic effort in selecting what to steal.

Like bad managers stealing credit for their subordinates' hard work, bad LLM advocates wish to normalize taking credit for the work the LLM does on their behalf.

Eating food doesn't count if you didn't grow it yourself.

"A game" is a different abstraction layer from "a piece of software", though. A game has designed mechanics, a scenario (level design, etc.), art/music assets, writing, and so on. I would say that if you're making all of those things yourself, but you're having an AI write the software that executes the game, then you're still "building a game" per se. "Developing a game" even.

Compare/contrast: people who develop games on top of high-level genre-specific game engines like RPG Maker are still considered to be "building a game." What's the difference between using a pre-made purpose-fit engine like RPG Maker, vs. asking an AI (or, for that matter, a contracted software company) to build you a custom purpose-fit engine?

I know now we all are somehow like 'product manager' when using agent/chatbot, and we just need to setup requirements and point out issues. But in game area, I am not so sure models/agent are capable enough, especially involves heavy graphic related development and human interaction things? Since from my personal experience, even codex and claude code not so stable and efficient when dealing visualisation things like figures, videos and even much easier one - web-browser.

They are a bit weird with game development at the moment.

They can one shot entire games, with relatively minor issues.

And obviously asking for small code snippets and integrating them yourself has been well supported for five years.

But in Agent mode... not so much. I was asking frontier models to make simple changes to my Pong game (you know like the one from 1972) and it constantly failed to make simple changes or would break something else in the process.

The main issue is that they can't see what they're doing. Actually one of the agents tried playing the pong game by screenshotting every frame, and it ran for about 20 minutes before I realized what it was doing, and told it to calm down.

It takes about 10 seconds to process an image, so it was running the game at 0.1 frames per second... 600x slower than realtime. The technology is not quite there yet.

If your game is something turn-based though, with discrete States and well-defined transitions between them, they can help out a lot more with that.

Depends a lot on what model you use. Claude Sonnet/Opus, Deepseek, Mimo, M3 are smart enough to figure out how to create debug views, add single-frame screenshot and testing harnesses.

Sounds like every other plagiarist.

Seems visulization is still too expensive and sluggish. I tried to build things like tetris and snake eater with fable5 and got-5.6 sol, 90% of build successful in one shot and remaining needs 1~2 rounds debug. The whole experience is much better than months earlier, where a snake eater takes me 3~4 manually compile/run and input the error message and abnormal manually again and again.

But I am curious about building a larger scale game with/without professional engines like Godot. I highly doubt the model/agent is intelligent enough to setup the entire development environment correctly, thus may be struggle to test and debug themselves. Maybe something like GUI-agent a future?

BTW, I also tried to make a entire LLM-driven bot to play boardgames like chess or mahjong, but it has the same latency issues since most of the online platforms have strict per-turn time limit like 15-30s, which is almost impossible for LLM/agent to react in time.

Just so everyone who reads this understands; despite the criticism in this comment, it's being incredibly charitable to LLMs in the context of making games. They're not particularly good at it and you wouldn't want most of the code they write in your game and especially not in the underpinnings of your game.

It's not true as a general rule that they could give you correct small code snippets five years ago; I had frontier models not even understanding enough of the OpenGL API to make separate framebuffers correctly. I asked this as a test and the resulting recommended code would of course compile, but leave someone who didn't already know the correct way to do it staring at a black screen and potentially spending the next hour debugging that.

If you think you're somehow going to make an actual game worthy of people's time with LLMs, you are almost certainly mistaken. Not only because it'll be a complete, unworkable mess with really bad fundamentals, but also because if you're vibecoding the actual technical part you probably also think game design is just about "having ideas" or something else that's really reductive and incorrect.

> Not only because it'll be a complete, unworkable mess with really bad fundamentals

Wait til you see the code for some of the best selling games of all time ;)

I've worked at two major studios and have seen some terrible code. Neither place had code as terrible as LLMs produce for games, they can't even get non-negotiable basics right a lot of the time.

"Is there telemetry? Yes, and it is opt-out: set ANTE_TELEMETRY=off"

Opt-out every time is unacceptable. If you opt out once it should be enough. One accidental execution path without the right environment and you're spewing telemetry to spaghetti knows where.

> One accidental execution path without the right environment and you're spewing telemetry to spaghetti knows where.

That would also be true with opt-in by environment variable, though.

True, it would be better to have a mandatory policy in a system location like /etc, would could over-ride environment if specified, and a command line option to show telemetry status. But accidental omission is more likely than accidental opt-in.

called out the most asked questions - where is the source - telemetry opt-in/opt-out in README of https://github.com/AntigmaLabs/ante

Considering that ripgrep, git, and, you know, other dev tools are part of the toolbox, then why ship them inside this executable? And, furthermore, if you ship them, then why stop there?

so it is more for being self contained and works out of box if being deployed in a bare linux environment. we tried shell out to `rg` it didn't work very well and instead spending time handling the args parsing and jugging string output, we decided to spend time on building Grep natively for agent. It is just a start~

as for why stop here yep, the goal is to be able to find the best sweet spot in being self contained v.s. all-in-one bloat ware. for example, we still use a bundled `tmux` skill for the orchestration.

so it is just llamafile with few additional application in bundle?

it is more like an actual harness with a managed and pinned llama the control is inverse.

Harness is compute and Model is data

Wow. Great concept!

> Is there telemetry? Yes, and it is opt-out

Fuck you. Will there ever be a decent agent where the answer to this is "No"?

it can be opt-out, opt-in, but the telemetry needs to be there to actually improve the agent.

the point is you can redirect all the telemetry to your own exporter with nothing leaves your machine and all the benefits.

this would go very hard with a lightweight gui

yes, the goal is to perfect the `ante serve` so it is easy to build gui. We are building one internally to test the protocol version

Hi HN, I'm Mohan from Antigma Labs. Ante is a coding agent that ships as one self-contained ~15MB binary: the TUI, an embedded ripgrep, local PDF/OCR, and a natively managed llama.cpp engine are all inside. No runtime dependencies, no node_modules, no account.

- Ante installs a pinned, checksum-verified official llama.cpp build matched to your machine (Metal on Apple silicon; CUDA, Vulkan, or CPU on Linux) and handles upgrades when the pin changes. - It discovers GGUF files already on disk (~/.ante/models, the llama.cpp and Hugging Face caches), attaches to llama servers already running on local ports, and estimates RAM/VRAM from model size and context window before anything loads. - `ante --offline-model /path/to/model.gguf "prompt"` boots the server, runs the session, and shuts it down. `/offline-mode` does the same interactively; `ante serve --offline-model` loads a model once for many clients. - No API key, no account. Once the model is on disk, inference needs no network at all; set ANTE_TELEMETRY=off and no telemetry is exported either.

On capability, we'd rather publish the number than oversell: we benchmark local models with the same harness and auditable runs as frontier ones, and Qwen3.6 27B (a 17 GB download) scores 56.2% on Terminal-Bench 2.1 across 445 trials (live results: https://antigma.ai/eval). That's a real gap from frontier models. The design bet is that you mix: hosted providers and local live in the same catalog, `/providers` switches mid-session, so sensitive repos or high-volume work go local and hard problems go frontier.

Hosted models work with your own keys or subscription. But nothing about trying Ante requires signing up for anything: download the binary, point it at a GGUF.

Offline mode is under active development and has rough edges with the overview at https://ante.run/local/overview. I'll be in the comments.

Not sure why this was dead but I vouched. It would be nice if telemetry was opt-in, otherwise this looks awesome and can't wait to try it!

will add those soon!

Opt-out telemetry is a hard no for me, sorry.

Right? How hard is it to just ask a single opt-in question during installation. Opt-out just seems lazy, especially if I have to dig though configs to get to it.

Agreed! When a tool is explicitely marketed for offline use, opt-out telemetry feels especially contradictory. Should Definitely be opt-in by default...

feedback received, it was carry over from the preview dev build.

Why not make it Actually Portable Executable using Cosmopolitan Libc, like Llamafile, to make it run on Windows/Linux/MacOS? Why don't you support Windows with CUDA?

even with power of AI, we are mere human and still slow. Adding this to backlog.

Thanks!

Where is the source code?

for now only some of core crates is migrated, will do so progressively

Many people have slow computers, but for agent it is no problem. Only run LLM slow too.

yes. Offline mode is choice, should be able to run frontier one first and then figure out how do incorporate local models as real workhorse

Here you say:

  > Ante installs a pinned, checksum-verified official llama.cpp

But in README:

  > Ante ships its own inference engine

May I suggest you use the first phrasing in both places. I took it as Ante devs had written their own engine and I doubt I'm the only one.

there is one toy version https://github.com/AntigmaLabs/nanochat-rs README updated.

the public repo README is updated.

Reply on news.ycombinator.com