tgpt — AI chatbots in the terminal
A cross-platform command-line program (Go, GPL-3.0) for talking to LLMs straight from a
shell prompt. You install one binary and run tgpt "your prompt"; it prints the answer.
No browser, no account for the default path. Runs on Linux, macOS, Windows, and FreeBSD,
with install scripts plus the usual package managers (Homebrew, Pacman, Scoop,
Chocolatey) and go install. Latest release 2.11.1 (Feb 2026); ~3.2k stars.
What it actually is
tgpt is a multi-provider chat frontend, not an agent. It sends a prompt to a backend and relays the reply — there is no tool use, no file editing, no multi-step planning, no autonomy loop. That makes it the minimal, non-agentic end of the terminal-AI tooling spectrum, and a useful boundary marker for this wiki: the same “AI in the terminal” niche that holds claw-code, oh-my-pi, and Gemini CLI also holds tools that are pure input→output multiplexers. (See synthesis for where this bounds the spoke’s subject.)
Its one distinctive axis is provider breadth rather than capability depth: one CLI fronting roughly ten backends, several of them free.
Supported backends (per the repo’s providers.md)
Text generation:
- openai — all models; needs an API key; supports custom/compatible endpoints (so it doubles as a generic OpenAI-compatible client).
- Gemini — free API keys; defaults to
gemini-2.0-flash. - deepseek — needs an API key.
- Groq — free key available; “many models.”
- MiniMax — needs a key; defaults to
MiniMax-M2.7. - Ollama — local models (the open-weight / on-device path; no key).
- Pollinations — free, “many free models.”
- KoboldAI — free (koboldcpp / a HF Space).
- Isou — free; Deepseek-chat combined with a SearXNG search step.
Image generation: Arta (free) and Pollinations (free).
Also has HTTP/SOCKS5 proxy support and a built-in self-update (-u).
Why this is here (and the cross-spoke note)
Routed from the hub (Telegram drop). It lands in agentic-tooling-wiki because it belongs to the terminal-AI tool ecosystem the domain claims — it just sits at the simple, non-agentic end. The runner-up spoke is llm-providers-wiki: tgpt’s real interest is as an access-layer artifact over the model market — a single client that normalizes paid APIs (openai, deepseek, MiniMax), free hosted tiers (Groq, Gemini, Pollinations), and local open-weight models (Ollama) behind one interface, with custom OpenAI-compatible endpoints as an escape hatch. That aggregation story (free + paid + open-weight access collapsed into one CLI) is the cross-spoke context; the provider market itself lives in llm-providers-wiki.