Antigma Labs launches Ante offline terminal coding agent

Antigma Labs has introduced Ante, a self-contained coding agent designed to run directly inside the terminal. Built entirely in Rust, the tool compiles into a single 15-megabyte binary with zero runtime dependencies, positioning itself as a lightweight alternative to tools like Claude Code and Codex. The startup designed the agent to run fully offline by embedding a pinned, managed version of llama.cpp, allowing developers to execute local GGUF models, such as Qwen3.5-9B-Q4_K_M.gguf, without requiring an internet connection, account, or API key.

To prove its capabilities, the developers continuously evaluate Ante on the Terminal-Bench 2.1 benchmark, which consists of 89 tasks tested over five trials each. In its latest public run, Ante achieved an 82.7 percent success rate—completing 368 out of 445 trials—using the open-weight DeepSeek V4 Flash 0731 model, which cost approximately $68 in inference fees. When compared to Claude Code across 20 parallel tasks in Docker, Ante demonstrated significant resource efficiency, utilizing roughly seven times less peak memory, nine times less average CPU, and five times less disk input-output.

The architecture of Ante embeds heavy operations like Git and Grep directly into its single process to prevent resource leaks. It supports multi-agent orchestration, Model Context Protocol, and persistent session memory, and it can connect to more than 12 cloud providers, including OpenAI and Anthropic. Currently in an alpha preview for macOS and Linux, the core harness is distributed as a prebuilt binary under temporary preview terms, while its software development kit, JSONL communication protocol, and evaluation pipeline are open-sourced under the Apache 2.0 license.