Why Coding Agents Need a Real Browser (and Playbooks That Replay Without an LLM)
Giving coding agents (Claude Code, Codex, Cursor, VS Code) a real browser hits three walls: fresh headless browsers have no logins, every rerun burns LLM tokens, and local wrappers lack sealed credentials, human takeover
Giving coding agents (Claude Code, Codex, Cursor, VS Code) a real browser hits three walls: fresh headless browsers have no logins, every rerun burns LLM tokens, and local wrappers lack sealed credentials, human takeover, or audit trails.
We built Oya Browser for both local coding agents and production fleets.
One-click local MCP (no account, no key)
curl -fsSL https://oyabrowser.com/install-desktop.sh | sh -s -- claude-code
Each agent session opens in its own tab at http://127.0.0.1:47321/mcp while you keep yours, inspects read_console and read_network only on its own tabs, and lets you take over for MFA or approval steps.
Record once, replay as standard Playwright code
from oya_browser import Oya
with Oya() as oya, oya.browser.start() as browser:
browser.goto("https://httpbin.org/forms/post")
browser.ask("Order a large pizza for {{name}}", data={"name": "Ada"})
browser.to_playbook("pizza")
browser.play("pizza", {"name": "Alan"})
In our replay benchmark, 9/10 sites replayed with no model and healing off. For healthcare teams, Oya Cloud and the enterprise edition add HIPAA mode, signed BAAs (oyabrowser.com/baa), and a 6-year hash-chained audit trail (desktop and community editions are not covered by the HIPAA offering).
Comparisons
Free desktop app: oyabrowser.com | GitHub
Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes ā full credit and traffic to the original publisher.