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

TL;DR Highlight

Ante is a single Rust binary coding agent that runs fully offline with local GGUF models.

Who Should Read

A developer who finds Claude Code or Codex heavy and model-locked, and wants to run local models or mix providers

Core Mechanics

One ~15MB binary is all you need
- It downloads at about 15MB compressed and expands to a single 34.1 MiB Rust executable.
- Grep and git run in the same process, so there are no runtime dependencies.
Offline with just a GGUF file
- Point the built-in llama.cpp at a GGUF file and it runs with no API key, account, or internet.
- Run it as `ante --offline-model <path> -p "..."`.
83.9% on Terminal-Bench 2.1
- With open-weight DeepSeek V4.1 Flash it passed 370 of 445 trials, at about $18 of inference.
- It used official leaderboard constraints: 89 tasks, 5 trials each.
Lighter than Claude Code
- On the same 20 parallel Docker tasks it used ~7× less peak memory than Claude Code.
- It also used ~9× less average CPU and ~5× less disk I/O.
One profile redefines the agent

Community Reactions

Several complained there is no source code
- Commenters said linking a GitHub repo that only has binary releases and no agent source is confusing.
- The author said they added answers on source location and telemetry to the README.
Complaints that telemetry is on by default
- Commenters said having to set ANTE_TELEMETRY=off every run is unacceptable; opting out once should persist.
Doubt about the memory-savings claim
- One commenter argued harnesses are simple loops that should use little memory even in Python or TypeScript, citing pi agent.
Author admits local models lag
- The author said Qwen3.6 27B (17GB) scores 56.2% on Terminal-Bench 2.1, a real gap from frontier models.
- The intended design is mixing: sensitive or high-volume work goes local, hard problems go to frontier models.

How to Apply

Try a local model run today
- Run the install script, then point `--offline-model` at a GGUF file you already have and give it a small edit task.
- Compare the result with a hosted agent like Claude Code and note the quality gap.
Split work between local and hosted models
- Use offline mode for repos where code leakage is a concern.
- Switch to a hosted model with `/providers` only for hard tasks.
Add headless diff review to CI
- Put `git diff | ante -p "review this for security issues"` in your pipeline.
- Check output quality on a local branch first, then move it to CI.
Build a minimal agent with a profile
- Run `cp curated/pi.settings.json ~/.ante/` and then `ante --profile pi` for a four-tool agent.
- Run the same task against the default profile to see the effect of a replaced system prompt.

Try It

copy & run

# Paste into a terminal (macOS/Linux, beta preview)

curl -fsSL https://ante.run/install.sh | bash

# Run offline with a local GGUF model ({{model_path}} is your GGUF file path)
ANTE_TELEMETRY=off ante --offline-model {{모델경로}} \
  -p "add error handling to src/main.rs"

# Next, compare with a hosted provider
git diff | ante -p "review this for security issues"

Terminology

HarnessThe runtime wrapper that lets an LLM call tools, edit files, and handle permissions. If the model is the engine, the harness is the chassis and controls.

GGUFThe model file format read by llama.cpp for local LLMs. It bundles a model into a single file.

llama.cppA C++ engine that runs LLMs on your own machine. Ante embeds a pinned, managed version of it.

Terminal-BenchA benchmark measuring how well an agent completes terminal tasks. Think of it as a practical exam for coding agents.

TelemetryUsage data that software sends back to its developer.