A 15 MB terminal agent built to run by the thousands

What it does

Ante is a ~15MB Rust binary that lives in your terminal and acts as a coding agent. It reads prompts, manipulates files, runs tools, and can operate as an interactive TUI, headless CLI, long-lived server, or chat gateway. The project treats the agent harness—not the model or the prompts—as the problem worth solving.

The interesting bit

Instead of wrapping a Python or Node.js runtime, Ante embeds its own Grep, git, and local inference engine into a single process. That obsession with self-containment is what lets it claim ~7× less peak memory and ~9× less average CPU than Claude Code across the same 20 parallel Docker tasks, and run fully offline against a local GGUF file.

Key highlights

• Public, reproducible benchmarks: it runs Terminal-Bench 2.1 continuously and publishes pinned builds with raw audit logs. With open-weight GLM 5.2, it scores 74.6% for a top-7 verified leaderboard slot.
• Zero vendor lock-in: bring your own API key or local model; 12+ providers are supported out of the box, and no account is required—not even with Antigma.
• Multi-modal deployment: interactive TUI, headless CLI (ante -p), server mode (ante serve), or Slack/Discord bot (ante gateway).
• Multi-agent orchestration: spawn sub-agents and coordinate tasks across decentralized or centralized architectures.
• Extensible via custom skills, MCP, and persistent memory across sessions.

Caveats

• Alpha software: the README warns of breaking changes, incomplete functionality, and macOS/Linux support only.
• The resource comparison claims are specific to a 20-task Docker benchmark against Claude Code; your mileage may vary in other workloads.

Verdict

Worth a look if you want a fast, offline-capable agent core you can own and replicate. Skip it if you need Windows support or a polished, feature-complete product today.