Authenticated browser agents
Reuse the login. Keep the password out of chat.
Most useful browser work sits behind a login. Sign in yourself while the agent is paused. It then carries on inside the signed-in session without ever being handed the password.
Source review: August 24, 2026
Three options
Choose where the login should live.
Each option puts signed-in state somewhere different. Pick the one whose owner and reach match the job.
Repeatable tests
Save a Playwright storage-state file
Best for test accounts and reproducible automation. Treat the file as a credential: Playwright warns it can hold cookies and headers that impersonate the account.
Playwright authentication guide ↗Exact current tab
Connect to your active browser
Best when the agent must see the exact page already open on your screen. The trade-off: it inherits everything your everyday browser profile can reach.
Chrome DevTools auto-connect ↗Durable agent work
Use a dedicated browser profile
Best when signed-in tabs should stay visible across agent tasks, away from your everyday Chrome. Hronaut owns the browser process; clients connect to it over loopback MCP.
Compare browser ownership models →The Hronaut workflow
The five-step login handoff.
- 1
Create an agent workspace
Ask the agent to create a fresh, named workspace. Each agent workspace has its own isolated Chromium profile.
- 2
Open the sign-in page
The agent navigates to the login page and stops before any secret is needed.
- 3
Pause MCP
Use the pause control beside MCP ready. New commands are rejected, while Hronaut, the workspace, and its tabs stay open.
- 4
Authenticate yourself
Type the password, approve MFA, solve the CAPTCHA, or use Hronaut's human-only saved-password control. Keep those values out of chat.
- 5
Resume inside the session
Resume MCP once the signed-in page is visible. The agent should inspect fresh state and reclaim write ownership before it acts.
The workspace keeps its cookies and site storage. After a disconnect or app restart, the task must resume with its workspace ID and private resume key.
Security boundary
An authenticated browser is delegated authority.
Once you sign in, a trusted agent may act with that account's permissions and read data visible to page scripts. Keep the blast radius small:
- Least privilege: use a dedicated low-privilege account when practical.
- One role per workspace: create separate workspaces for separate roles.
- Local trust: enable Hronaut MCP authentication when other local processes are not equally trusted.
- Human moments: pause MCP before any sensitive manual action.
- Clean up: when the work ends, sign out or clear that website's cookies and storage if the session should not stay reusable. Hronaut's site-data controls apply only to the website you choose and do not clear unrelated workspaces.
Why a separate browser helps
The browser does not need to be a child of every agent session.
A current Codex issue documents globally configured headed browser MCP servers starting once per coding session and piling up. A separately launched browser with a stable loopback HTTP endpoint changes that: the app owns one browser, and Codex, Claude Code, Cursor, Copilot, OpenCode, or another client can disconnect and reconnect. Read the full task, transport, and browser lifecycle architecture →
Observed demand
Authentication without credential exposure
A current Claude Code issue asks for an agent to inherit a session authenticated out of band, instead of handling a password in plaintext.
Read issue #82668 ↗Observed demand
Browser lifecycle independent of the session
A current Codex issue describes one headed MCP browser process being retained by every live coding session.
Read issue #21984 ↗Limits and sources
Use the smallest browser authority that fits the job.
Hronaut is designed for interactive, visible browser work. It does not turn a human login into a safe unattended credential broker, and it is not a replacement for isolated multi-browser CI.
Try the dedicated-browser pattern
Keep the signed-in workspace. Change the agent task.
Install Hronaut, connect one MCP client, and try the handoff with a low-risk test account first.