AntigmaLabs/ante

Ante – A lightweight terminal‑based coding agent

What it is – Ante is a single‑binary, self‑contained coding agent that lives in your shell. It talks to LLM providers (OpenAI, Anthropic, Gemini, Grok, DeepSeek, etc.) or runs a local GGUF model via a bundled llama.cpp engine, so you can work completely offline. The binary is written in Rust, about 15 MB compressed (≈34 MiB unpacked), and has no external runtime dependencies.

Key capabilities

Feature	How it works
Interactive TUI	ante launches a terminal UI where you can type natural‑language prompts, see streamed LLM replies, and let the agent edit files, run commands, or browse the repo.
Headless / script mode	ante -p "<prompt>" runs a one‑shot task, ideal for CI, git hooks, or quick fixes.
Server mode	ante serve starts a daemon exposing a JSONL protocol over stdio, Unix socket or WebSocket, enabling editor plugins and custom front‑ends.
Gateway	ante gateway runs a Slack/Discord bot that forwards messages to the same agent core.
Multi‑agent orchestration	The /term command opens named tmux panes that run other agents (Claude, Codex, or another Ante instance) side‑by‑side, letting you watch and intervene in their work.
Profiles	A JSON settings file (.ante/settings.json or a named profile via --profile) defines system prompt, enabled tools, memory limits, and other behaviour. Switching profiles turns the same binary into a full‑featured assistant, a minimal “pi” agent, or a read‑only planner.
Tool integration	Built‑in tools include grep, git, file read/write, Bash execution, and a generic “tool” interface that can be extended.
Offline inference	Point ante at a local GGUF model (--offline-model path/to/model.gguf) and the bundled llama.cpp runs the model locally—no API key, no internet.
Extensible provider catalog	~/.ante/catalog.json lets you add arbitrary OpenAI‑compatible endpoints, custom auth headers, or extra request fields, so any LLM service can be used.

Performance claims – Benchmarked on the public Terminal‑Bench 2.1 suite (89 tasks, 5 trials each). With the DeepSeek V4.1 Flash model Ante achieved 83.9 % success, using roughly 7× less peak memory, 9× less CPU, and 5× less disk I/O than Claude Code in the same Docker workload.

Typical workflow

# install the binary (one‑liner)
curl -fsSL https://ante.run/install.sh | bash

# interactive coding session
ante

# quick one‑off fix
ante -p "find and fix the failing test in src/auth"

# run offline with a local model
ante --offline-model ~/.ante/models/Qwen3.5-9B-Q4_K_M.gguf \
     -p "add error handling to src/main.rs"

Ecosystem – The repo contains:

• ante-acp – an Agent Client Protocol server for IDE integration.
• ante-gateway – Slack/Discord bridge.
• Rust crates (ante-sdk, protocol‑shape, exec, llm) that expose the JSONL protocol and core primitives.
• ante-harbor – the adapter used for the public benchmark results.
• curated/ – ready‑made profile JSON files (e.g., pi, plan).
• examples/ – sample harnesses and custom integrations.

Who it’s for – Developers who want a fast, reproducible coding assistant that can run locally, be scripted, or be embedded in tools without pulling in heavyweight runtimes (Node, Python). The design emphasises tiny footprint, verifiable builds, and the ability to spin up thousands of identical agents.

License – Apache‑2.0 for the source code; the pre‑built binary is free for commercial use under a separate “Binary Preview Terms” agreement.

---

All details above are taken directly from the repository’s README.