Best browser automation MCP server for ChatGPT (Codex), tested

Dated results

• Playwright MCP passed in ChatGPT (Codex) on 2026-10-21 (Codex CLI 0.162.0-alpha.17.2, @playwright/mcp 0.0.83, connect time not recorded). Full result
• Chrome DevTools MCP passed in ChatGPT (Codex) on 2026-10-20 (Codex CLI 0.162.0-alpha.17.2, chrome-devtools-mcp 1.11.0, connect time not recorded). Full result
• Firefox DevTools MCP passed in ChatGPT (Codex) on 2026-10-23 (Codex CLI 0.162.0-alpha.17.2, @mozilla/firefox-devtools-mcp 0.10.6, connect time not recorded). Full result
• Agent Infra Browser (ByteDance) passed in ChatGPT (Codex) on 2026-10-14 (Codex CLI 0.162.0-alpha.17.2, package version not recorded, connect time not recorded). Full result
• Browser Use (MCP mode) failed in ChatGPT (Codex) on 2026-10-14 (Codex CLI 0.162.0-alpha.17.2, package version not recorded, connect time not recorded). Full result

Why Playwright MCP is the candidate

Playwright MCP passed every step in ChatGPT (Codex). It is the group’s site-wide candidate (judged 2026-10-21): Opened our test page, clicked 'Reveal code' and read back the code in both Claude Code and Cursor, with 4 browser calls in each. Chrome DevTools tied on every measure (passed both, 4 calls each); Playwright gets the nod for its smaller tool list (25 vs 30). ByteDance's agent-infra browser also passed both (11 calls); Browser Use failed both (Claude Code read the code only from a screenshot, and its browser didn't start under Cursor). On Oct 14 Playwright, Chrome DevTools and agent-infra also passed in Claude Desktop and ChatGPT (Codex); Browser Use failed in both (its Chrome didn't start). Confirmed Oct 21, 2026: a second run passed again in all 4 clients (Claude Code, Cursor and ChatGPT (Codex) on Oct 21 AM, Claude Desktop on Oct 21 PM; @playwright/mcp 0.0.83).

Provisional: based on one dated run per app (no repeat runs yet), so treat it as a candidate, not a final pick. How we test.

What failed or is partial in ChatGPT (Codex)

• Browser Use (MCP mode): failed on 2026-10-14. Recorded reason: answer missing ['TP-4242']; no tool result containing ['TP-4242']

Install in ChatGPT (Codex)

The exact snippet we tested is on the Playwright MCP page (pick the ChatGPT (Codex) tab). Not connecting? ChatGPT (Codex) MCP troubleshooting.

Not tested in ChatGPT (Codex) yet

Browser MCP, BrowserStack MCP Server, Browser Harness, Ui.Vision RPA (MCP), axuse, MCPSafari, Mobile MCP, Windows-MCP, Peekaboo, BrowserTools MCP. Listed from their source; no result in ChatGPT (Codex) yet.

Questions

What is the best browser automation MCP server for ChatGPT (Codex)?

In ToolPassed's hands-on tests in ChatGPT (Codex), Playwright MCP is our provisional candidate for browser automation: it passed install, connect, tools and an agent task on 2026-10-21. We ran 5 browser automation MCP servers in ChatGPT (Codex); 4 passed every step; not a pass: Browser Use (MCP mode) (failed).

Does Playwright MCP work with ChatGPT (Codex)?

Playwright MCP passed in ChatGPT (Codex) on 2026-10-21 (Codex CLI 0.162.0-alpha.17.2, @playwright/mcp 0.0.83, connect time not recorded).

Does Browser Use (MCP mode) work with ChatGPT (Codex)?

No, not in our test: Browser Use (MCP mode) failed in ChatGPT (Codex) on 2026-10-14 (Codex CLI 0.162.0-alpha.17.2, package version not recorded, connect time not recorded). Recorded reason: answer missing ['TP-4242']; no tool result containing ['TP-4242']