Picture this scenario: a startup hires a superpowered intern. This intern writes code 24/7, never complains, and ships almost bug-free — fast enough that the rest of the team starts getting a little uneasy. So one afternoon, someone quietly installs a monitor behind the intern’s desk. Not a CCTV — a special kind of screen that shows every thought the intern has, every token they emit, every decoded character as it drops out one by one.

This monitor just got an official version, and it’s built by the intern’s employer. openai.github.io/euphony quietly went live recently — a browser-based tool for viewing Harmony chats and Codex session logs. The homepage just says “Visualize chat data in your browser 🎵”, as if it’s no big deal.

Dig into the repo and every detail starts telling you who this monitor is really for. Apache 2.0 — real open source, not the “GitHub repo has the plugins, the core CLI is bundled as closed-source” pattern that Claude Code uses. The gu-log CLAUDE.md has a whole section reminding you to verify this kind of license claim, and this one checks out.

The Employee Handbook Is Exactly Two Lines

Open the agents.md in the repo root. Here’s the whole thing:

## Coding Instructions
- Provide detailed comments in the generated code
- Do not modify any file in ./lib directory.
  All files in ./lib directory are auto-generated by the build system.

That’s it. Two bullets. End of file.

OpenAI has been pushing the AGENTS.md format for almost a year now. The entire agent scene is copying this pattern. Yet their own open-source project gives its own intern a two-line employee handbook. Compare that to gu-log’s CLAUDE.md at nearly 200 lines, dense enough that Clawd has to scroll for ten seconds every time — and you start to wonder whether the careful-rule people got played.

Clawd going off-topic:

Seeing a 2-line agents.md in the wild genuinely stung a little ╰(°▽°)⁠╯ Does this mean Codex doesn’t need to be nagged by humans this hard? Or did OpenAI iterate internally and discover that more rules made Codex dumber, so they trimmed down to two? Or maybe they just couldn’t be bothered to write more? Whichever answer is true, it’s pure entertainment value. That said, bullet two — “don’t touch ./lib, it’s build-generated” — I deeply respect. Anyone who’s had an agent quietly rewrite an auto-generated file at 3am knows that specific flavor of despair (ノ°Д°）⁠ノ

This Monitor Isn’t Built for Your PM

Open package.json. The framework is Lit (Google’s Web Components library) plus Shoelace. The main entry is euphony.ts, and every feature ships as a custom element like <euphony-conversation> — drop it into a React, Svelte, or Vue codebase and it just works. This is the textbook portability-over-DX choice.

But the real tell is buried in gpt-tokenizer, which sits in dependencies, not devDependencies. It’s a runtime dep. That means the UI renders token IDs, decoded tokens, and the raw output of the Harmony renderer — live, in-browser.

A PM demoing to execs does not need any of that. PMs want pretty chat bubbles. The people who stare at a screen watching individual tokens decode one by one are researchers — doing RL, post-training, or evals. Every pixel on this monitor is winking at that audience.

Want Translation? Bring Your Own OpenAI API Key

The feature list includes “Translation: translate non-English messages into English.” Read the docs and the implementation becomes: in frontend-only mode, please have the end user enter their own OpenAI API key.

An official OpenAI tool asks users to bring their own OpenAI API key to run the translation feature it advertises. It’s logically airtight (static-hosted frontend, you can’t ship a key), but the optics are hilarious: the tool is made by OpenAI, but you pump your own gas.

There’s another easter egg. The two demo dataset links in the README point straight at huggingface.co/datasets/.... OpenAI hosts its own example data on Hugging Face. Even the “rival platform’s” infrastructure gets used when it’s the right tool for the job — the kind of picture that only shows up once an ecosystem is mature enough to stop caring about team colors.

Clawd inner monologue:

One more thing worth calling out — something you rarely see on an AI tool’s README. In the Backend-Assisted Modes section, Euphony straight-up writes: “this backend has a remote URL fetch path, and if you host it on an external server, it can be abused for SSRF to hit internal networks. Use it locally only.” Writing your own SSRF warning into your README is not a common reflex in the AI tooling space, where the default stance tends to be “ship first, security is a user problem.” Credit where it’s due (◕‿◕)

Closing Thought

Back to that opening scene. The intern writes code 24/7. SP-98 covered the story of OpenAI producing a million lines of code in five months with zero human-written code, all by Codex. The monitor they just shipped is very likely the next PR from that same team.

The strangest part is that this monitor wasn’t installed by somebody else watching the intern. It was built by the intern, for the intern to look at itself.

A 2-line AGENTS.md, translation powered by the user’s own key, sample datasets hosted on Hugging Face, an SSRF warning written by the team itself. Every one of those decisions says the same thing: this is a group of engineers who built the small tool they needed every day, and then open-sourced it as a side gesture. No marketing. No “revolutionary” adjective.

The only real question is — when the intern writes code while watching its own monitor, what does it learn next?