Ante is a terminal coding agent shipped as a single Rust executable with no runtime dependencies, downloaded at about 15 MB compressed. Heavy parts such as grep and git are embedded in the one binary and process rather than shelled out, and local inference is handled by a managed build of llama.cpp. It is in beta preview on macOS and Linux, with WSL suggested on Windows.
The project treats the harness rather than a co-trained model as its subject, and evaluates Ante across model families instead of one flagship. Terminal-Bench 2.1 runs continuously under the official leaderboard constraints of 89 tasks and 5 trials each; the latest full run reports 82.7% with the open-weight DeepSeek V4 Flash 0731 across 368 of 445 trials on Ante 0.preview.71, at about $68 of inference, and each result pins the exact build and links the raw run. Across the same 20 parallel tasks in Docker, Ante is reported to use about 7 times less peak memory, 9 times less average CPU, and 5 times less disk I/O than Claude Code.
Point Ante at a GGUF file and the whole loop runs locally with no API key or network. For hosted models it ships 17 maintained provider presets, including Anthropic, OpenAI, Google Gemini, Grok, DeepSeek, and Open Router, each handling that provider's wire dialect, key or OAuth flow, and thinking and streaming behaviour. Anything else is one entry in ~/.ante/catalog.json with a wire style from the four Ante speaks, an auth style, and extra headers or body fields.
Behaviour lives in a settings file, and --profile swaps the whole file per run, covering system prompt, tool set, skills, and memory, with anything omitted falling back to defaults and explicit flags still winning. A curated pi profile is the minimal end of that range at four tools and a short replacement prompt, and a built-in bare profile drops skills, MCP servers, session saving, and auto-memory.
The repository holds the documentation site, the protocol and SDK crates behind ante serve, the adapter used to reproduce benchmark runs, and curated profiles and skills. The core harness itself is developed privately during the alpha and ships as a prebuilt binary, with crates moved into the repository as they stabilise.
Features
- Single binary: a self-contained Rust executable with zero runtime dependencies and embedded grep and git
- Native offline mode: a managed llama.cpp engine runs GGUF models locally, handling engine installation, model discovery, and memory management
- Four run modes: interactive TUI, headless one-shot prompts, ante serve over a JSONL protocol for editors, and ante gateway for Slack or Discord
- Provider presets: 17 maintained presets plus a catalog file for any other endpoint across four wire styles and bearer, header, query, or OAuth auth
- Profiles: --profile swaps the whole settings file, including the system prompt, tools, skills, and memory
- Public evaluation: continuous Terminal-Bench 2.1 runs pinned to downloadable builds, with raw runs linked for audit
- Protocol and SDK: open protocol-shape and Rust SDK crates for building clients against the daemon
- Session control: resumable saved sessions, interrupt-and-steer during a turn, and file or folder context by mention
