agentty-org/agentty

AI pair programming in your terminal — one static binary, sub-ms startup, any model

What is agentty?

agentty is a native, single‑binary coding assistant that runs in your terminal. It lets you talk to a large language model (Claude, GPT, Grok, Ollama, etc.) to write, edit, and run code, while automatically retrieving the most relevant parts of your repository so the model only sees what it needs. The whole thing is written in C++26, ships as a ~16 MB static executable, and works on Linux, macOS, Windows, and even Android/Termux.

Core ideas

Idea	What it means for you
Zero‑runtime dependencies	No Node, Python, or `npm install`. Just download the binary (or build it) and run it.
Selective retrieval (RAG)	A built‑in hybrid BM25 + dense‑embedding engine fetches only the code snippets that matter, cutting the prompt context by 80 %+ compared with dumping the whole repo.
Model‑agnostic	Works with any OpenAI‑compatible endpoint, Claude, Groq, DeepSeek, Gemini, local Ollama models, etc. Switch providers on the fly with `^P`.
Sandboxed execution	Shell/ build commands are run inside OS‑enforced sandboxes (seccomp/landlock on Linux, seatbelt on macOS) so the assistant can’t accidentally damage your system.
Air‑gapped mode	When the target machine has no internet, agentty can relay traffic over an SSH‑tunneled SOCKS5 connection, keeping the session isolated.
Smart Mode	An orchestrator routes easy turns to cheap models and reserves a flagship model for planning‑heavy tasks, automatically learning the right effort level for each repository.

Main features (as described in the README)

Instant start‑up – cold start under 1 ms, keystroke‑to‑pixel latency ~1 ms.
Full tool suite – read/write/edit, git, grep, search_docs, search_code, web search, task management, and more, all exposed through a TUI.
Local RAG engine – hybrid BM25 + HNSW vector search, reciprocal‑rank fusion, pseudo‑relevance feedback, MMR diversification, GraphRAG expansion, and optional generative rerankers, all implemented in‑process C++.
Sandboxed commands – Linux uses seccomp + landlock + cgroup2 (fallback bubblewrap); macOS uses seatbelt + rlimits (fallback sandbox‑exec).
Air‑gapped SSH relay – agentty airgap user@host sets up a TLS‑pinned SOCKS5 tunnel.
CLI modes – agentty run "fix the failing test" for one‑shot scripts, --json/--events for automation, --read‑only/--no‑network for safe CI runs.
Key bindings – Vim‑style shortcuts for sending, navigating threads, opening the provider picker, toggling Smart Mode, looping prompts, etc.
Integration with Zed – speaks the Agent Client Protocol, so you can use it as a Claude‑Code replacement inside the Zed editor.

Getting started (quick‑start)

# Install (Linux/macOS/Windows/Termux) – the script picks the right package
curl -fsSL https://raw.githubusercontent.com/1ay1/agentty/master/install.sh | sh

cd /path/to/your/project
agentty               # first run opens an auth prompt for your API key

You can also install via package managers:

Debian/Ubuntu: dpkg -i agentty_<version>_amd64.deb
Fedora/RHEL: dnf install https://…/agentty-<ver>-1.x86_64.rpm
macOS: brew tap 1ay1/tap && brew install agentty
Windows: scoop bucket add 1ay1 … && scoop install agentty or winget install agentty.agentty
Build from source with a C++26 compiler (git clone … && cmake -B build && cmake --build build).

Typical workflow

Open a repo – cd my‑app && agentty.
Ask a question – type something like Add a unit test for the login handler and hit Enter.
Agent retrieves – the RAG engine pulls the most relevant code snippets and shows them as a quiet 📚 card.
Model replies – the LLM proposes edits, maybe with fenced code blocks.
Run or apply – press ^G to execute a suggested command, ^R to review a diff, or ^S to let Smart Mode plan a multi‑step refactor.
Iterate – continue chatting; the tool keeps a thread history, and you can switch providers at any time with ^P.

Who might benefit?

Developers who prefer the terminal over heavyweight IDE extensions.
Teams with strict security – sandboxed execution and air‑gapped mode keep secrets safe.
People using many LLM providers – the same binary works with Claude, OpenAI, Groq, local Ollama, etc.
Open‑source contributors – the project is MIT‑licensed, fully static, and can be compiled for any platform.

Where to learn more

Documentation – https://agentty.org/docs (CLI, sandbox, retrieval, Smart Mode, keybindings).
Comparisons – https://agentty.org/compare/agentty-vs-claude-code, vs Aider, vs Cursor.
Community – Discord, IRC, and the GitHub issue tracker.

In short: agentty is a fast, self‑contained terminal coding assistant that brings modern LLM‑powered code generation, context‑aware retrieval, and secure sandboxed execution to any OS, without the overhead of Node/Python ecosystems.