1ay1/agentty

What it solves

agentty is a high-performance terminal-based coding agent designed to replace bulky runtime-dependent tools. It addresses the problem of slow startup times, high token costs associated with "context dumping" (sending entire repositories to an LLM), and vendor lock-in by providing a fast, model-agnostic binary that retrieves only the most relevant code snippets.

How it works

Written in C++26, agentty operates as a single static binary with no dependencies on Node.js or Python. It uses a sophisticated local RAG (Retrieval-Augmented Generation) engine to fetch specific code slices using a hybrid of BM25 keyword search and dense embeddings.

Key architectural features include:

• Smart Mode: An orchestrator-worker design where a flagship model plans and delegates tasks to cheaper models based on complexity.
• Sandboxing: Shell and build commands are executed in OS-enforced sandboxes (seccomp/landlock on Linux, seatbelt on macOS) to ensure safety.
• Local RAG Pipeline: A multi-stage funnel involving Reciprocal Rank Fusion, re-ranking, and GraphRAG expansion to provide high-precision context without cloud dependencies.

Who it’s for

Developers who want a blazing-fast, privacy-focused coding assistant in their terminal that can work with any LLM provider (including local models via Ollama) and supports air-gapped environments.

Highlights

• Extreme Performance: Cold start under 1ms; single static binary (~16.7 MB).
• Model Agnostic: Supports Claude, GPT, Groq, OpenRouter, Ollama, DeepSeek, xAI, Gemini, and Fireworks.
• Advanced Retrieval: Local hybrid BM25 + embeddings with GraphRAG and semantic search over source code.
• Security First: Default sandboxing for shell calls and an air-gapped mode via SSH relay.
• Integrated Tooling: Full suite of tools for reading, writing, editing, and searching code, plus the ability to run code blocks interactively in the terminal.