Antigma Labs releases offline coding agent Ante as single Rust binary
New self-contained agent harness operates without external dependencies or API keys, targeting local inference and scalable agent architectures.
Antigma Labs has released Ante, a self-contained coding agent distributed as a single 15MB binary written in Rust. The tool is designed to operate offline without external dependencies or API keys, utilising a built-in llama.cpp engine to support local inference via GGUF models. Available for macOS and Linux, the software positions itself as a resource-efficient alternative to existing agents such as Claude Code or Codex.
The developers claim that Ante utilises approximately seven times less peak memory, nine times less average CPU, and five times less disk I/O than Claude Code across 20 parallel tasks in Docker. The project is currently in an alpha preview phase, with the core harness developed in a private repository and shipped as a prebuilt executable. While the core libraries are open-sourced under the Apache 2.0 license, the prebuilt binary is provided under specific binary preview terms that permit commercial use during the alpha period.
In continuous runs of Terminal-Bench 2.1 under official leaderboard constraints, Ante achieved an 82.7 per cent success rate using the open-weight DeepSeek V4 Flash 0731 model. The benchmark involved 89 tasks with five trials each, costing approximately $68 in inference fees. Antigma Labs noted that DeepSeek reported the same 82.7 per cent success rate for the model when measured with its own unreleased harness in minimal mode.
Ante supports more than 12 providers out of the box via environment variables or OAuth, and allows for custom provider configuration through a catalog file. The software includes a server mode that runs the agent as a long-lived daemon communicating over a structured JSONL protocol, making it suitable for editor plugins and web user interfaces. The project philosophy emphasises a cellular-native approach, aiming for tiny, expendable, and massively replicated agents.
The tool is described as a self-contained agent harness that self-organises within the terminal. It includes features such as multi-agents, skills, Model Context Protocol (MCP) support, and persistent memory. Antigma Labs states that the architecture is designed to handle millions of agents self-organising and communicating at massive scale, where runtime optimization becomes economically significant. Full documentation is available at docs.antigma.ai.
