A terminal that talks back to multiple LLMs

What it does

gpt-cli is a Python-based terminal client for chatting with large language models. It wraps OpenAI, Anthropic, Google Gemini, Cohere, and OpenAI-compatible APIs (Together, OpenRouter, local models via LM Studio) behind a single gpt command. You can run interactive sessions, fire one-off prompts, or even generate and edit shell commands before executing them.

The interesting bit

The project treats “assistants” as configurable personas rather than just model switches. You define them in a YAML file—pirate, dev, bash expert, whatever—each with its own system prompts, model, temperature, and even API endpoint overrides. The --execute flag for the bash assistant is a nice touch: it generates a command, drops you into $EDITOR to sanity-check it, then runs it. Claude 3.7’s extended thinking mode is also exposed via a --thinking budget flag.

Key highlights

• Supports 5+ LLM providers through a unified interface
• Custom assistants with per-persona model parameters and system prompts
• One-off prompt mode (--prompt) for scripting and piping
• Token count and price tracking built in
• !include directive to load file contents into assistant context
• Multi-line input mode and markdown toggle for readability

Caveats

• The README’s installation section is Linux/macOS-centric; Windows users are on their own
• No mention of conversation history persistence—sessions appear ephemeral
• The uvx one-liner in “Try now” requires uv, which isn’t discussed elsewhere

Verdict

Worth a look if you live in a terminal and juggle multiple LLM providers. Skip it if you need persistent chat history, rich UI features, or a batteries-included Windows experience.