November 7 agent guide: OpenCode Browse for signed-in browser agents

November 7, 2026 was a quiet frontier-model Saturday and a loud browser-agent day. OpenCode Browse pairs an MCP server with a Chromium extension so coding agents can drive the browser the user is already signed into. Bacancy Technology covers it as infrastructure for computer-use and browser-use agents: real sessions, human handoffs, allowlists, and capture, not a new foundation model.

What shipped on November 7

OpenCode Browse: the signed-in browser as an MCP tool

Browse gives agents a browse MCP server. They open and drive tabs with Playwright, take screenshots, capture network traffic as a redacted HAR, and record the screen, usually in background tabs grouped out of the way. Supported browsers include Chrome, Edge, Brave, Arc, Helium, Vivaldi, Opera, and Chromium. Install paths include the Chrome Web Store extension plus npx @opencode/browse install or the documented curl / PowerShell installers.

Handoffs and session controls

• Human handoff: login, 2FA, passkeys, and payments stay with the user; the agent continues after Continue.
• Lend a tab: the toolbar button hands an open tab to an agent or takes it back.
• Named sessions survive restarts, each with its own tabs.
• Allowlists limit a session to named sites (session_allowlist / allowedOrigins).
• Secrets helpers can type stored credentials on allowed sites without exposing values in transcripts or artifacts.

Why it matters for agentic systems

Most browser agents fail on authentication walls or force teams to store credentials in agent memory. Browse inverts that: the agent borrows the human’s authenticated context under local relay control (documented around 127.0.0.1) and returns the tab when identity steps appear. That design is powerful and risky. Bacancy Technology’s rule is the same as for any computer-use agent: dedicated browser profile, tight allowlists, no banking or payroll tabs in the agent group, and recorded sessions for audit.

How this fits the November 4–6 window

November 4 opened Chinese open-frontier weights and APIs. November 5 widened Cursor’s path to hosted internal apps and narrowed OpenAI cyber access. November 6 hardened Claude Code MCP tasking and clarified agent memory. November 7 supplies the browser surface those agents increasingly need. Read the four days as one control-plane sequence, not four unrelated launches.

Builder checklist

1. 01
   Install on a dedicated profile

   Use a non-primary Chrome profile, pair the extension, and confirm the local relay before connecting Claude Code or Cursor.
2. 02
   Set allowlists before first real task

   Create a session limited to the staging domains the agent needs. Refuse broad *.com allowlists.
3. 03
   Practice one handoff

   Run a login that requires 2FA, complete Continue, and verify the agent resumes without receiving the password in the transcript.

Conclusion

November 7’s high-value agent story is OpenCode Browse: MCP-native control of the signed-in browser with human handoffs and allowlists. It is not a model release. It is the missing browser lane for teams already shipping coding and computer-use agents.

Related reading

• November 6 Claude Code 2.1.319 and memory: /blog/november-06-claude-code-2-1-319-agent-memory-2026
• November 5 GPT-6 Cyber and Live Apps: /blog/november-05-gpt6-cyber-cursor-live-apps-2026
• November 1 Mage-VL-Cowork computer use: /blog/november-01-mage-vl-cowork-computer-use-2026
• MCP topic map: /topics/mcp
• Agentic AI topic map: /topics/agentic-ai

Need browser-agent controls designed for production?

Bacancy Technology designs MCP allowlists, computer-use sandboxes, and browser-agent runbooks for enterprise teams. Consulting: /consulting/ . Contact: /contact/ .

Sources: opencode.ai — browse ; GitHub ; Chrome Web Store listing for browse (updated November 7, 2026).