gptty-web 0.1.0

Terminal client for existing ChatGPT web sessions, powered by chatgpt-web-adapter.

gptty

English version. Russian version: README.ru.md

Terminal client for existing ChatGPT web sessions.

[!WARNING] Not the official OpenAI API. Uses an existing ChatGPT web session. Web backend behavior may change.

gptty is the successor to webchat-openai-cli. The project is being migrated from a standalone script into a terminal-native product powered by chatgpt-web-adapter.

The package distribution name is gptty-web because the PyPI name gptty is already occupied. The installed command is still gptty.

Product Direction

SDK = chatgpt-web-adapter
CLI = gptty

gptty is intended for terminal workflows:

gptty chat
gptty auth status
gptty auth refresh --mode wait
gptty ask "explain this error"
gptty ask --image screenshot.png "describe this UI"
git diff | gptty ask "review this patch"
gptty attach https://chatgpt.com/c/...
gptty send "continue from here"
gptty messages --last 5 --format markdown
gptty status --format json
gptty export --format markdown --output conversation.md

gptty ask, gptty send, the default gptty chat path, conversation inspection commands, and conversation export are SDK-backed. The legacy interactive runtime remains available through gptty chat --legacy while feature parity is migrated in later PRs.

Current Features

• minimal SDK-backed interactive chat through gptty chat
• inspect auth data through gptty auth status
• refresh auth_data.json through gptty auth refresh
• attach existing conversations through gptty attach
• send prompts to attached, explicit, or new conversations through gptty send
• SDK-backed image prompts through gptty ask --image and gptty send --image
• inspect attached or explicit conversations through gptty messages and gptty status
• export attached or explicit conversations through gptty export
• output modes for messages, status, send, and export: plain, json, markdown
• legacy interactive chat fallback through gptty chat --legacy
• one-shot SDK-backed prompts through gptty ask
• centralized stdin policy for pipe-friendly prompts
• pipe-friendly prompts, for example git diff | gptty ask "review this patch"
• streaming replies in the terminal
• minimal SDK chat state file: gptty_state.json
• legacy state file for --legacy: webchat_state.json
• atomic writes for local state and auth_data.json
• legacy image prompts through /img in gptty chat --legacy
• auto and wait auth capture modes
• English and Russian CLI localization in the legacy runtime

Requirements

• Python 3.10+
• system curl available in PATH
• Chrome or Chromium for auth capture
• valid auth_data.json for an existing ChatGPT web session

Installation

Base install:

python -m pip install gptty-web

Install with browser auth-capture dependencies:

python -m pip install "gptty-web[auth]"

From checkout for development:

python -m venv .venv
python -m pip install --upgrade pip
python -m pip install -e ".[auth,test]"

On Windows cmd.exe:

python -m venv venv
venv\Scripts\activate.bat
python -m pip install --upgrade pip
python -m pip install -e ".[auth,test]"

Get auth_data.json

Check the current auth file without opening a browser:

gptty auth status

Use JSON output for scripts:

gptty auth status --format json

Refresh auth data through the CLI wrapper:

gptty auth refresh --mode wait

Fast mode for an already logged-in browser session:

gptty auth refresh --mode auto

In wait mode the browser stays open until the chat is ready. After that, send any message manually in the browser to trigger auth capture.

Optional: override the one-shot probe prompt used by auto mode:

gptty auth refresh --mode auto --probe-prompt "Ping"

The legacy script entrypoints remain available from a checkout:

venv\Scripts\python.exe auth_fetcher.py --mode wait
venv\Scripts\python.exe auth_fetcher_wait.py

After a successful capture, auth_data.json will appear in the project directory. See docs/auth.md for lifecycle details and troubleshooting notes.

Run the CLI

Attach an existing ChatGPT conversation:

gptty attach https://chatgpt.com/c/...

Send a prompt to the attached conversation:

gptty send "continue from here"

Pipe stdin into the attached conversation:

git diff | gptty send "review this patch"

Send to an explicit conversation without changing first through attach:

gptty send --to https://chatgpt.com/c/... "continue there"

Start a new conversation and store its returned conversation reference in gptty_state.json:

gptty send --new "start a new conversation"

Send image prompts through the SDK-backed commands:

gptty ask --image screenshot.png "describe this UI"
gptty ask --image https://example.com/chart.png "summarize this chart"
gptty send --image diagram.webp "continue with this image"
gptty send --image before.png --image after.png "compare these images"

--image accepts local file paths, http(s) URLs, and data URIs. It can be used more than once. Supported SDK image formats are PNG, JPEG/JPG, GIF, and WebP.

Inspect the attached conversation:

gptty messages --last 5
gptty status

Use JSON or Markdown output for scripts and exports:

gptty messages --last 5 --format json
gptty messages --last 5 --format markdown
gptty status --format json
gptty send --format json "summarize the current thread"
gptty export --format markdown --output conversation.md
gptty export --format json --output conversation.json

When gptty send uses --format json or --format markdown, streaming is disabled internally so the output stays complete and parseable.

You can also inspect or export an explicit conversation without attaching it:

gptty messages https://chatgpt.com/c/... --last 5
gptty status https://chatgpt.com/c/...
gptty export https://chatgpt.com/c/... --last 20 --output conversation.md

gptty export defaults to Markdown output. When --output points to an existing file, add --overwrite to replace it.

Minimal SDK-backed interactive chat:

gptty chat

The SDK-backed chat loop currently supports:

/help
/new
/exit
/quit

Run the full legacy interactive runtime:

gptty chat --legacy

One-shot SDK-backed prompt:

gptty ask "explain this error"

Pipe stdin into the prompt:

git diff | gptty ask "review this patch"

When stdin and a prompt are both present, gptty ask and gptty send send stdin as context, followed by the prompt under User prompt:.

Force reading stdin:

gptty ask --stdin "summarize this input"

Ignore piped stdin:

cat noisy.log | gptty ask --no-stdin "explain this from the prompt only"

Disable streaming and print the final response:

gptty ask --no-stream "summarize this session"
gptty send --no-stream "summarize this conversation"

Legacy entrypoint, still supported from a checkout:

python main.py

You can also override local paths:

gptty auth status --auth ./auth_data.json
gptty auth refresh --auth ./auth_data.json --mode wait
gptty attach https://chatgpt.com/c/... --auth ./auth_data.json --state ./gptty_state.json
gptty send --auth ./auth_data.json --state ./gptty_state.json "hello"
gptty send --auth ./auth_data.json --state ./gptty_state.json --image ./screenshot.png "describe this"
gptty export --auth ./auth_data.json --state ./gptty_state.json --output conversation.md
gptty chat --auth ./auth_data.json --state ./gptty_state.json
gptty chat --legacy --auth ./auth_data.json --state ./webchat_state.json
gptty ask --auth ./auth_data.json --timeout 120 "hello"

Useful Legacy Chat Commands

Available in gptty chat --legacy:

• /help
• /models
• /new
• /list
• /use <chat_id>
• /reset
• /img <path_or_url> :: <prompt>
• /settings
• /model <name>
• /lang <en|ru>
• /ws <true|false>
• /effort <standard|extended|off>
• /metrics <true|false>

Important Files

• auth_data.json - local auth data, do not commit it
• gptty_state.json - minimal SDK-backed chat state, do not commit it
• webchat_state.json - legacy chat history and runtime settings, do not commit it

Notes

• auth_data.json is the primary auth source.
• ChatGPT web-session auth may expire after some time; in practice, expect to refresh it periodically.
• Run gptty auth status when requests start failing or before long terminal sessions.
• .env is optional. If present, accessToken is used as a fallback even when auth_data.json is missing, but a full auth_data.json remains the most compatible setup.
• In auto mode, auth refresh sends one probe message to trigger capture. The default text is "Hello", and you can override it with --probe-prompt.
• In wait mode, auth refresh does not send the probe automatically. Log in or register, then send any message manually in the browser to trigger capture.
• New auth captures write canonical accessToken plus the legacy-compatible api_key field.
• Do not mix cookies and api_key/accessToken from different accounts.
• Local state and auth files are written atomically to reduce the chance of truncated JSON after interruption.
• If main.py says that curl is missing, install system curl.exe and check curl --version.

Troubleshooting

• curl not found Install system curl.exe and make sure curl --version works.
• auth_data.json is missing Run gptty auth refresh --mode wait, complete login in the browser, then send any message in the chat window.
• Auth may be expired Run gptty auth status. If it reports expired, run gptty auth refresh --mode wait.
• gptty auth refresh says auth dependencies are missing Reinstall auth dependencies with python -m pip install -e ".[auth]" from checkout, or python -m pip install "gptty-web[auth]" from an installed package.
• gptty send, gptty messages, gptty status, or gptty export says there is no attached conversation Run gptty attach <url-or-id> first, pass a conversation URL/id directly to the command, or use gptty send --new.
• gptty ask --image or gptty send --image says an image file does not exist Check the local path, or pass an http(s) image URL instead.
• ImportError: cannot import name 'nodriver' Reinstall auth dependencies with python -m pip install -e ".[auth]". Recent g4f releases use zendriver instead of the older nodriver package name.
• The wrong account opens in auth refresh The browser profile already contains another session. Log out there first, or use the wait mode and sign in to the intended account.
• Requests start failing after working before Your session cookies or api_key/accessToken may have expired. Regenerate auth_data.json with gptty auth refresh --mode wait.
• gptty chat starts but cannot answer Check that auth_data.json exists and the captured browser session still belongs to the same account.
• gptty chat --legacy starts but cannot answer Check that auth_data.json exists, curl is installed, and the captured browser session still belongs to the same account.

Status

This repository is in transition from webchat-openai-cli to gptty.

PR0 establishes the package skeleton and console command. PR1 adds the SDK client boundary. PR2 adds the first SDK-backed command, gptty ask. PR3 centralizes stdin pipe handling. PR4 migrates the default gptty chat path to a minimal SDK-backed loop with legacy fallback. PR5 adds attach/messages/status conversation operations. PR6 adds send-to-attached, explicit, and new conversation workflows. PR7 adds shared output modes for messages/status/send. PR8 adds conversation export. PR9 adds SDK-backed image prompts for ask/send. PR10 adds auth status/refresh UX. PR11 prepares the first gptty-web release flow. Later PRs will add richer pipe workflows and SDK chat /img parity.

See CHANGELOG.md and docs/release.md for release details.