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.
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.
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."
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.
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 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.
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.
> 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.
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.
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.
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?
Agreed! When a tool is explicitely marketed for offline use, opt-out telemetry feels especially contradictory. Should Definitely be opt-in by default...
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.
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.
"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."
One could even argue what defines AI instructability is heuristics as opposed to specifics
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 coding agent is more like a carpenter, a mason, an electrician,... rather than a hammer in that case.
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.
- 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.