agentty

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_asserts — 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.