Terminal-first AI coding agents let you stay on the command line — no IDE switch — while the model reads files, edits code, and runs commands. Here's what to look for, and where agentty fits.
What makes a great terminal coding agent
- Fast startup — you invoke it constantly; a slow cold start is death by a thousand cuts.
- Small footprint — ideally a single binary, no runtime to install.
- Provider choice — Claude, OpenAI, Groq, local models — not locked to one vendor.
- Retrieval — sends only relevant code, not your whole repo.
- Safety — commands sandboxed by default.
- Open source — auditable and self-hostable.
Why agentty
agentty was built to hit all of the above:
- Native C++26 single binary — 16.7 MB, ~3 ms cold start, no Node/Python.
- Multi-provider — Claude Pro/Max, OpenAI, Groq, OpenRouter, Cerebras, or local Ollama.
- Built-in hybrid retrieval with a learning loop.
- Sandboxed by default, SSH air-gap in one command.
- Runs in Zed over ACP; speaks MCP.
- Open source (MIT).
Other options
Depending on your workflow, Aider (mature git flow), OpenCode (open TUI), and vendor CLIs (Claude Code, Codex CLI, Gemini CLI) are all worth knowing. See our honest comparisons and alternatives.
Install in one line
curl -fsSL https://agentty.org/install.sh | sh