Terminal Chat (OpenAI from your shell)

Tiny bash script to send one-off prompts to OpenAI and stream the reply directly in the terminal. No conversation history; every call is stateless.

Requirements

• Bash, curl
• jq
• OpenAI API key set as OPENAI_API_KEY

Install

1. Create a user bin dir and copy the script there: mkdir -p "$HOME/.local/bin" && cp chat.sh "$HOME/.local/bin/chat.sh".
2. Make it executable: chmod +x "$HOME/.local/bin/chat.sh".
3. Ensure your shell can find it (zsh): add export PATH="$HOME/.local/bin:$PATH" to ~/.zprofile.
4. (zsh) Add an alias that disables globbing so * and ? pass through: alias chat='noglob "$HOME/.local/bin/chat.sh"'.

Update later with a single command (no git checkout needed):

mkdir -p "$HOME/.local/bin" && curl -fsSLo "$HOME/.local/bin/chat.sh" https://raw.githubusercontent.com/dillweed/terminal_chat/main/chat.sh && chmod +x "$HOME/.local/bin/chat.sh"

Usage

• Inline: chat "Prompt here"
• Interactive single line: run chat, type at >, press Return.
• Multiline / piped:
  • Run chat, press Return on the blank prompt, type text, finish with END on its own line.
  • Or pipe/redirect: cat notes.txt | chat

Tips

• If you skip the alias, wrap prompts that contain shell metacharacters (* ? [ ]) in quotes.
• When piping, pass zero CLI arguments; stdin is ignored if args are present. To include an instruction, put it as the first line of stdin (see examples).

Configuration

• OPENAI_CHAT_MODEL (default gpt-5-codex)
• System message is defined near the bottom of the script; edit to taste.

What it prints/keeps

• Streams the model output as it arrives.
• Saves the last response to /tmp/chat_last_output.txt.
• On API errors, writes the raw payload to /tmp/chat_error.json.

Troubleshooting

• "OPENAI_API_KEY is not set" → export it in your shell rc file.
• "jq is not installed" → brew install jq (macOS) or apt-get install jq.
• Nothing printed? Ensure the model name is valid and network egress to api.openai.com is allowed.

Examples

• Inline: chat "What's the rsync syntax to mirror directories on a remote server and log differences without making changes?"
• Interactive single line: chat → at the prompt type Give me a one-line ffmpeg to downsample audio to 64k mono → Return
• Interactive multiline: chat → press Return on the blank prompt → paste or type multiple lines, e.g.:

  I need a Bash loop that retries curl up to 5 times with exponential backoff.
  Show me the loop and explain the backoff math.
  END
  

• Piped file with instruction (stdin only):

  { echo "Explain what this script does and point out any obvious bugs."; cat script.sh; } | chat
  

Changelog (recent)

• Switched to streaming Responses API; removed spinner.
• Cleaned prompts/alias instructions; no stray backslash lines.
• Stopped rewriting model output (no automatic bullet insertion).