We turned codex into an in-browser AI agent!

Resolved 💬 2 comments Opened Oct 20, 2025 by irichard00 Closed Nov 5, 2025

💡 Likely answer: A maintainer (etraut-openai, contributor) responded on this thread — see the highlighted reply below.

Hi OpenAI Team,

First, thank you for creating such a beautiful implementation with Codex — it has been incredibly inspiring for us. At AI Republic, we’ve extended its potential by turning it into a fully in-browser AI agent that can operate web pages for general tasks beyond coding:
👉 https://github.com/The-AI-Republic/browserx

We’ve named our project “Codex For Chrome” (also known as BrowserX) in honor of the original Codex project, and we clearly credit OpenAI in the project’s README file.

Currently, the project is adopting the same Apache 2.0 license to maintain consistency with the upstream Codex repository. We’re writing to ask whether we may continue using the name “Codex For Chrome”, or if you prefer that we change it before making it a fully open-source release.

Thank you for your time and guidance!

Warm regards,
Richard
Founder, AI Republic

View original on GitHub ↗

2 Comments

kordless · 10 months ago

README admits the whole project is “derived from OpenAI’s open-source Codex reference implementation,” yet the same document warns that no one may create or publish derivative works during their “alpha” phase (README.md:17, README.md:25). Under Apache 2.0 that restriction isn’t enforceable; they’re trying to bolt proprietary terms onto existing Apache code.
package.json still declares the repo’s license as ISC, even though the README promises a future Apache 2.0 release once they resolve trademark questions (package.json:18, README.md:33). That mismatch makes the current licensing status murky while they tell others not to redistribute.
The core runtime is marked as a direct port of codex.rs structures—CodexAgent, Session, TurnContext, etc.—but the README’s “not open source yet” language suggests they want a proprietary claim on those same Apache-licensed internals (src/core/CodexAgent.ts:1, src/core/Session.ts:1, src/core/TurnContext.ts:1).
The background service worker wraps those ported components and exposes the same submission/event queue model Codex already published (src/background/service-worker.ts:26). Nothing novel there—just the upstream design transplanted into a Chrome extension.
Tool registration is lifted straight from Codex’s tool suite; they even keep the feature flags from the default config, flipping them on/off with the same names (src/tools/index.ts:31, src/config/defaults.ts:64).
They call their API-key handling “encryption,” but it’s just reverse + base64, identical to Codex’s simple obfuscation (src/utils/encryption.ts:16). Claiming stronger privacy protection while reusing that insecure method is misleading.
Overall footprint: core+tools+models add up to tens of thousands of lines that mirror Codex logic (46k lines excluding tests). They’re effectively re-bundling Apache code, relabeling it, and telling the community to keep out until they bless a release.

etraut-openai contributor · 9 months ago

It's great to see all of the innovative ways that the community is using codex! Thanks for sharing.