Ante — a compact AI agent in Rust that runs directly in the terminal and supports local models

If you've used console AI assistants like Claude Code, you've probably noticed how resource-hungry they are. Under the hood, they typically run Node.js or Python environments, and processes quickly fill up RAM. When you need to run a single agent for a local task, you can tolerate this. But if you launch multiple parallel sessions, the system starts to noticeably slow down.

Engineers from Antigma Labs tried to solve this problem radically. They wrote Ante — an agent as a single compiled Rust binary of about 15 MB.

Why another console agent

The project developers started from a simple idea. The agent should be fast and consume minimal system resources. In Ante, code search and Git utilities are built directly into the binary. No calls to external interpreters or extra processes occur during file searches.

The tool operates in four modes:

• Interactive TUI for everyday console work
• One-off execution via the ante -p "задача" command for scripts and CI
• Server mode ante serve that communicates via the JSONL protocol and is suitable for writing code editor plugins
• Network gateway ante gateway for connecting a bot to Slack or Discord

Resource efficiency and benchmarks

The authors measured resource consumption during execution of 20 parallel tasks inside Docker. Comparison with Claude Code showed a significant gap: Ante consumes roughly 7 times less peak RAM, 9 times less CPU resources, and 5 times less disk I/O.

On the Terminal-Bench 2.1 benchmark (89 tasks with 5 attempts each), the fresh Ante build achieves 82.7% successfully solved tests with the DeepSeek V4 Flash model. Running all tests costs approximately $68 in API requests. All run results are published publicly for audit.

Autonomous operation without internet and API keys

Particularly interesting is the built-in engine based on llama.cpp. If you can't send source code to external servers due to security requirements or simply have no network access, you feed the agent a regular GGUF format file.

The command for local launch looks like this:

ante --offline-model ~/.ante/models/Qwen3.5-9B-Q4_K_M.gguf \
  -p "добавь обработку ошибок в src/main.rs"

In this mode, all logic and code generation execute entirely on your machine. No accounts or API keys are required.

If you prefer cloud models, the tool supports 17 ready-made presets: Anthropic (Claude Sonnet/Opus), OpenAI, Google Gemini, Grok, and OpenRouter. Third-party proxies connect through the configuration file ~/.ante/catalog.json.

Profiles and task-specific configuration

The utility's behavior changes flexibly through profiles. All configuration is stored in a JSON file that defines the system prompt, available tools, and session memory.

For example, the pi profile trims the toolset down to four basic actions: reading, writing, editing, and executing commands in Bash. For file searches, rg is used; for web requests, curl; and additional tasks are delegated through sub-agents.

cp curated/profiles/pi.settings.json ~/.ante/
ante --profile pi

What's with the source code

There's a nuance here that the authors honestly warn about. The project is currently in the alpha preview stage. The core library and main binary are shipped in compiled form. However, the source code for the SDK (crates/agent-sdk), interaction protocol (crates/protocol-shape), evaluation system (ante-harbor), and documentation is open under the Apache 2.0 license.

Developers are gradually releasing core components publicly as they stabilize (for example, the process launch module crates/exec is already open). Discussion of open core code is happening in issue #21 on GitHub.

Quick start and usage examples

Installation takes a couple of seconds since there should be no external dependencies in the system:

curl -fsSL https://ante.run/install.sh | bash

Examples of everyday commands:

# Поиск и исправление падающего теста
ante -p "найди и исправь падающий тест в src/auth"

# Проверка изменений перед коммитом
git diff | ante -p "проверь эти изменения на уязвимости"

# Запуск с указанием конкретного провайдера и модели
ante --provider openai --model gpt-5.5 -p "перепиши модуль работы с базой данных"

Who should try it

Ante will be useful for developers who enjoy working in the terminal but are tired of the sluggishness of other AI assistants. The project is great if you want to run agents locally without sending code to the cloud or plan to embed an AI agent into your automation and CI/CD scripts without extra overhead.