# agentty

> A blazing-fast, native C++26 terminal coding agent — a drop-in alternative to claude-code shipped as a single 13.6 MB static binary with sandboxing, multi-model support, and sub-millisecond startup.

agentty is an open-source, native C++26 terminal coding agent created by Ayush Bhat and released under the MIT license. It positions itself as a drop-in alternative to `claude-code`, shipping as a single 13.6 MB static binary with no Node, Python, Electron, or `npm install` required. The project is pre-1.0 and actively developed, with daily builds for Linux, macOS, and Windows.

## What It Is

agentty is a command-line AI pair-programming tool that runs a full agentic coding loop — reading files, writing edits, running shell commands, searching codebases, and browsing the web — directly in your terminal. It signs in with an existing Claude Pro/Max OAuth subscription or an API key, and can also be pointed at OpenAI, Groq, OpenRouter, Together, Cerebras, or a local Ollama model. The core design is a pure-functional update loop (`(Model, Msg) → (Model, Cmd)`) rendered by a sister TUI engine called maya, with the Anthropic provider speaking HTTP/2 + SSE directly through an in-house nghttp2 + OpenSSL stack.

## Architecture and Performance

The project's headline differentiator is its native C++26 implementation versus the Node.js runtime that powers the official `claude-code` client. According to the project's own benchmarks on a Linux x86_64 machine:

- Cold-start `--help`: ~2 ms (agentty) vs ~150 ms (claude-code)
- `--version`: ~2 ms vs ~60 ms
- Binary on disk: 13.6 MB vs 222 MB + Node runtime
- No V8 GC pauses mid-stream

Subprocesses use `posix_spawn` + `poll(2)` with in-process kill deadlines. File writes are atomic (`write` + `fsync` + `rename`). The permission matrix is a `constexpr` with `static_assert`s — changing a policy cell breaks the build rather than a test.

## Key Capabilities

- **Sandbox by default**: Every shell and build call runs inside `bwrap` (Linux) or `sandbox-exec` (macOS). `~/.ssh`, `/etc`, and other projects are read-only even for approved bash calls.
- **One-command SSH air-gap**: `agentty airgap user@host` relays bytes over SOCKS5-over-SSH with TLS pinned end-to-end on real upstreams.
- **Agent Skills and memory**: Drop a `SKILL.md` under `.agentty/skills/` to teach conventions; `remember`/`forget` provide durable cross-session memory. The project cites research (arxiv.org/abs/2410.03981) claiming agent accuracy on codebases with internal DSLs jumps from ~20% to ~85% with curated skills.
- **Local RAG engine**: Hybrid BM25 + embeddings with reranking over docs, skills, and memory.
- **Isolated subagents**: The `task` tool spawns a subagent with its own context window for self-contained jobs.
- **MCP, both ways**: Serve agentty's tools to any MCP client with `mcp-serve`, or consume other MCP servers via `.agentty/mcp.json`.
- **Zed ACP integration**: `agentty acp` speaks the Agent Client Protocol, making agentty a first-class agent panel in the Zed editor.
- **Run code blocks**: `^G` runs fenced shell commands interactively on the real terminal — sudo prompts work, output streams live.

## Multi-Model Support

agentty defaults to Claude via OAuth (Pro/Max) or API key, but supports live model switching mid-thread with `^P`. Supported providers include OpenAI (GPT and o-series), Groq (Llama/Mixtral on LPUs), OpenRouter, Together AI, Cerebras (wafer-scale inference), and local Ollama models at `localhost:11434`. Any raw OpenAI-compatible endpoint is also accepted via `--provider host:port`.

## Update: v0.2.8

The latest release is v0.2.8, published on July 16, 2026. The repository was last pushed on July 18, 2026, indicating active development. The project is explicitly pre-1.0 and described as "moving fast." Prebuilt release binaries ship for Linux (x86_64, aarch64) and Windows (x86_64); macOS builds from source via Homebrew tap. Package distribution covers Debian/Ubuntu `.deb`, Fedora/RHEL `.rpm`, Arch AUR (`agentty-bin`), Alpine `.apk`, Scoop, and winget — all auto-submitted by GitHub Actions on each tagged release.

## Features
- Single 13.6 MB static binary with zero runtime dependencies
- Sub-millisecond cold start (~2 ms on Linux x86_64)
- Sandbox by default via bwrap (Linux) / sandbox-exec (macOS)
- One-command SSH air-gap mode over SOCKS5-over-SSH
- Multi-model support: Claude, GPT, Groq, OpenRouter, Together, Cerebras, Ollama
- Live model switching mid-thread with ^P
- Agent Skills via SKILL.md for codebase conventions
- Durable cross-session memory with remember/forget
- Local RAG engine (hybrid BM25 + embeddings, reranked)
- Isolated subagents via task tool
- MCP server and client support
- Zed editor integration via Agent Client Protocol (ACP)
- Run code blocks interactively from AI replies (^G)
- Persistent threads with automatic compaction
- Permission profiles: Ask, Write, Minimal
- File mentions (@), symbol jumps (#), command palette (/)
- Image paste support (PNG, JPEG, GIF, WebP)
- Git tools: status, diff, log, commit rendered natively
- Web fetch and web search tools
- Workspace boundary enforcement

## Integrations
Claude (Anthropic) via OAuth Pro/Max or API key, OpenAI (GPT and o-series), Groq (Llama, Mixtral), OpenRouter, Together AI, Cerebras, Ollama (local models), Any OpenAI-compatible endpoint, Zed editor (Agent Client Protocol), MCP (Model Context Protocol) servers and clients, bwrap (Linux sandboxing), sandbox-exec (macOS sandboxing), SOCKS5-over-SSH (air-gap relay)

## Platforms
WINDOWS, MACOS, LINUX, WEB, API, CLI

## Pricing
Open Source

## Version
v0.2.8

## Links
- Website: https://agentty.org
- Documentation: https://agentty.org/docs/
- Repository: https://github.com/1ay1/agentty
- EveryDev.ai: https://www.everydev.ai/tools/agentty