An interactive coding shell powered by ChatGPT or Gemini via browser automation. No API keys needed — uses your existing subscriptions.
I wanted a coding agent like Claude Code or GitHub Copilot — something that could read files, run commands, and fix code autonomously in the terminal.
The problem:
- Claude Code and Codex are expensive. API pricing at $3–15 per million tokens adds up fast with heavy daily use. The subscription plans ($100–200/month) are cheaper but still significant.
- They have strict rate limits. Hit the limit mid-task and you're blocked.
- Running models locally sounds great but sucks in practice. Local models (even 32B) are noticeably worse at coding than GPT-4o or Gemini. You need a good GPU (~$800+), and even then token speed is slow.
The insight: most people already pay for ChatGPT Plus or Gemini Advanced ($20/month each) and those subscriptions have generous limits on the web interface — but no way to use them programmatically without paying again for API access.
rawagent bridges that gap. It drives a real Chrome browser so it looks like normal web usage, and exposes the LLM as a tool-calling coding agent — giving you a Claude Code-like experience at no extra cost.
- Python 3.10+
- Google Chrome (installed)
- A ChatGPT or Gemini account (free or paid)
git clone https://github.com/yourname/browser-llm-agent
cd browser-llm-agent
pip install -e .
playwright install chromium# use Gemini (recommended — free)
rawagent --llm gemini
# use ChatGPT
rawagent --llm chatgpt
# open both, start on Gemini
rawagent --llm bothOn first launch a Chrome window opens. Log in to ChatGPT or Gemini manually, then press Enter in the terminal to start.
Type naturally — just like Claude Code:
you [gemini]> create a fastapi app with a /health endpoint
you [gemini]> add input validation to the POST /users route
you [gemini]> run the tests and fix any failures
| Command | Action |
|---|---|
/new |
Start a fresh conversation (clears LLM memory) |
/switch |
Swap between ChatGPT and Gemini |
/help |
Show help |
/quit |
Exit |
Paste multi-line code or text directly into the prompt — it is sent as a single message.
The agent can use these tools automatically:
| Category | Tools |
|---|---|
| Workspace | workspace_snapshot, project_detect, read_many_files, apply_patch |
| Files | read_file, write_file, edit_file, multi_edit, regex_edit, replace_lines, insert_at, append_file, delete_file, move_file, list_dir, make_dir |
| Search | glob, grep, find_files |
| Shell | bash, bash_status, bash_kill, python_exec, node_exec |
| Web/browser | web_fetch, browser_navigate, browser_screenshot, browser_click, browser_fill, browser_get_text, browser_eval |
| Git | git_status, git_diff, git_log, git_commit, git_branch, git_stash |
| Todos | plan_create, todo_add, todo_list, todo_update, todo_delete |
| Memory | memory_save, memory_get, memory_list, memory_search, memory_delete |
The system prompt is Codex-style: inspect the worktree before edits, preserve user changes, plan multi-step work, make targeted patches, run relevant verification, and report changed files plus any skipped checks.
The agent can remember things across sessions:
you [gemini]> remember that this project uses PostgreSQL and FastAPI
you [gemini]> what stack does this project use?
you [gemini]> add a todo: write tests for auth module
you [gemini]> show all todos
you [gemini]> mark todo abc123 as done
Todos and memory are stored in ~/.llm-agent/.
| File | Contents |
|---|---|
~/.llm-agent/memory.db |
Persistent memory (SQLite) |
~/.llm-agent/todos.json |
Todo list |
~/.llm-agent/history |
Input history (arrow keys) |
VS Code terminal not opening — make sure the command name does not conflict with another tool in your shell. rawagent should be safe.
Tool calls not executing — the agent parses JSON from the LLM response. If the LLM changes its output format, restart with /new and try again.
Selectors broken after a site update — edit src/browser_llm_agent/llm/chatgpt.py or gemini.py and update the CSS selectors to match the current page structure.