Agentty: a coding agent in one 16.7 MB binary

Most terminal coding agents arrive as an npm package and bring a runtime with them. Agentty takes the other route: one static executable, written in C++26, released under the MIT license. Its repository presents it as an alternative to the established command-line agents.

Delivery is the differentiator

The README puts the binary at 16.7 MB and describes it as having no runtime dependencies — no Node, no Python, no npm install. That choice shows up in how the project is distributed. Listed install paths include a curl script, packages for Debian and Ubuntu, Fedora and RHEL, Arch via the AUR and Alpine, plus Homebrew on macOS, Scoop and Winget on Windows, and a native build for Termux on Android.

Building from source is the exception to the low-friction story. It requires a C++26 toolchain — GCC 14 or newer, a recent Clang, or MSVC — which is a stricter requirement than most developer machines carry by default.

The speed numbers are the project's own

Two figures anchor the pitch: roughly 3 ms to cold start, and about 1 ms from keystroke to pixel. Both come from the project's own documentation. We did not benchmark them, and the repository page does not state the hardware or the measurement method behind them.

The same caveat applies to the popularity signals. At the time of retrieval the page showed 597 stars, 17 forks, 1,767 commits on the master branch and no open issues. No version number or release date was visible on the fetched page, and the only maintainer identified is the account 1ay1.

No provider lock-in, and a cost cascade

Agentty is not tied to one vendor. It lists Anthropic, OpenAI, Groq, Ollama, OpenRouter, DeepSeek, xAI and Google Gemini, along with any OpenAI-compatible endpoint. The first launch prompts for an API key; a local Ollama model, the README notes, needs none at all.

The economics show up in Smart Mode, an orchestrator-workers design that splits work across three tiers: a flagship model for strategy, a mid-tier model for implementation, and a cheap one for utility work. The project says it learns per workspace, remembering the effort level the cascade settled on and the task decompositions that worked. It toggles off with a keystroke and is described as a byte-for-byte no-op when disabled.

Sandboxing, offline use and skills

Shell calls are said to run inside bwrap on Linux and sandbox-exec on macOS, with file tools refusing paths outside the workspace. The README names one gap itself: on unrooted Android, kernel limits mean shell and build tools run unsandboxed. For locked-down environments there is an air-gapped mode over SSH with TLS pinning.

Code search combines BM25 with dense embeddings, and the semantic half optionally leans on a local Ollama server. The agent also loads Agent Skills from a skills directory, and it can run inside the Zed editor through the Agent Client Protocol.

What we could not check

This report rests on a single source — the project's own repository page. Every performance, size and capability figure here is therefore a vendor claim rather than a verified measurement. The startup timings, the strength of the sandbox, and how the agent holds up against commercial alternatives on longer tasks all remain unverified.