OpenAI has given Codex direct access to the Chrome DevTools Protocol, moving its coding agent from static code generation into live browser manipulation. Reported by TestingCatalog on June 16, the feature lets Codex inspect console output, network traffic, page payloads, and rendered DOM state, then rewrite that DOM in real time.
The practical scope is wider than it may sound. A coding agent that can see a live page, read its JavaScript performance profile, and modify its markup is doing something categorically different from one that writes code to a file. It can recolor a theme, adjust spacing via annotations, pull structured data from a rendered page, or detect a runtime error that only surfaces in the browser. That closes a loop that has historically required a human developer to shuttle information between the browser panel and the editor.
Prior to this, wiring Codex or other agents to browser inspection tools meant routing through external connectors and third-party plumbing. Bringing CDP access in-house, paired with Codex’s own embedded browser, lets OpenAI control the data layer and build proprietary tooling on top of it. That is not a convenience change; it is a platform move.
The feature is opt-in under Settings and organizations can disable it with a toggle. OpenAI has excluded the EEA, the UK, and Switzerland from the initial rollout. That geographic carveout follows a now-familiar pattern: features involving data signals from live environments tend to get held back from jurisdictions with active AI and privacy enforcement. The exclusion signals that OpenAI views CDP access as data-sensitive enough to warrant caution, even in a developer-facing tool.
In practice, the current implementation is rough. TestingCatalog notes it runs slowly, degrades under pressure, and sometimes requires a restart. The models appear undertrained on the CDP toolset; results come through, but typically after careful prompting and multiple attempts. The technology works. The product is not yet reliable.
The CDP rollout fits inside a larger infrastructure strategy. Days before this release, OpenAI announced the acquisition of Ona (formerly Gitpod) to give Codex persistent cloud environments for long-running tasks. That acquisition addresses a different constraint: agents that can browse and modify pages in a session still need a durable workspace if the task runs for hours. CDP access answers the question of what the agent can perceive. Persistent cloud environments answer the question of how long it can keep working.
Codex now serves more than five million weekly users, according to OpenAI. At that scale, a live-browser capability, even an early one, starts shaping how developers think about the boundary between writing code and operating a website directly through an agent. The long-term direction being tested here is an AI layer that sits in front of the web and adapts what a user sees. Faster models and infrastructure capable of supporting that at scale do not exist yet, but the protocol-level plumbing is going in now.
Teams evaluating Codex for front-end workflows should treat CDP access as a capability to prototype against, not ship on: the model quality and performance headroom are not there yet, but the architectural integration suggests it will improve faster than a third-party connector ever could.
Source reporting by TestingCatalog (Alexey Shabanov), published June 16, 2026.
