# What you will learn

• What dtzp555-max/ocp actually does: an OpenAI-compatible shim in front of your Claude Pro/Max login
• How to install it, issue LAN keys, and point Cursor, Continue, or any OpenAI client at it
• Why the repo's "disciplined cli.js alignment" matters more than the cache or quota features
• Where it falls short: one maintainer, young code, and a subscription you do not own
• Whether the $0 pitch survives contact with Anthropic's terms of service

# Why would anyone want an OpenAI-compatible Claude?

You already pay for Claude Pro or Max. You also have a pile of tools — Cursor, Continue, Cline, a homegrown script on your laptop, a tablet on the couch — that all expect the OpenAI /v1/chat/completions shape. Historically you had two options: pay twice, or write a glue layer yourself. OCP is that glue layer, packaged.

Per the repo's own description, it turns your subscription into an OpenAI-compatible endpoint for IDEs and devices, with LAN authentication, per-key quotas, a response cache, and what the author calls "disciplined cli.js alignment." That last phrase is the tell. The project is not a clean-room reimplementation of Claude's API. It is a wrapper that keeps pace with Anthropic's command-line client as that client changes.

That is the whole ballgame, and it is why I am writing about a repo that is still small enough to read in an afternoon. The interesting question is not whether you can get an OpenAI-shaped endpoint out of a Claude subscription. The interesting question is whether the bridge holds when the thing it bridges moves.

# How do you install OCP and put it on your LAN?

The project is JavaScript, so the path of least resistance is a clone and a Node run. The exact commands live in the repo README, and you should read them there rather than trust a third party — including me — to transcribe them. The shape looks like this:

git clone https://github.com/dtzp555-max/ocp
cd ocp
npm install
# authenticate against your Claude Pro/Max subscription
# then start the local server, binding it to your LAN interface
npm start -- --host 0.0.0.0 --port 8787

Once it is up, you configure your client the way you would configure it for OpenAI, swapping the base URL:

# in your IDE or client settings
OPENAI_BASE_URL=http://192.168.1.50:8787/v1
OPENAI_API_KEY=ocp_your_issued_key

The LAN auth and per-key quota features are the reason this is more than a one-line proxy. You are not exposing your subscription to the whole house with a single shared secret; you can mint a key per device and cap each one. The response cache is a nice touch for the repetitive prompts that IDEs love to resend — the same context prefix, the same system prompt, over and over.

If you have been following the broader argument that AI codes better than most devs — and that was never the job, you can see where this fits: the bottleneck is not model quality inside the editor, it is plumbing and cost. OCP attacks the plumbing.

# What does "disciplined cli.js alignment" really mean?

This is the phrase to interrogate. Anthropic's own tooling — the CLI that ships as cli.js when you install the Claude Code package — is the reference implementation for how a Pro/Max subscription is exercised programmatically. OCP does not call a public API with an API key. It sits behind the same door the CLI uses.

That has two consequences. First, it is the source of the $0 claim: you are not paying per token, you are riding a flat-rate subscription. Second, it is the source of the fragility: every time Anthropic changes how the CLI authenticates, batches requests, or interprets headers, OCP has to chase it. "Disciplined alignment" is the maintainer's promise to do that chasing promptly. It is a maintenance commitment disguised as a feature bullet.

I have seen this movie before. Projects that piggyback on a first-party client tend to be excellent for months and then break on a Tuesday. The ones that survive are the ones whose maintainers treat upstream changes as the primary work, not the side quest. Whether dtzp555-max has that stamina is unknowable from a repo this young. What you can evaluate today is the code: is the alignment logic isolated, tested, and easy to patch? If yes, you can live with the risk. If it is smeared across the proxy, walk away.

# Who is this actually for, and where does it hurt?

OCP is for the solo developer or small team that already pays for Claude Pro/Max, runs a home lab or office LAN, and wants IDE autocomplete and chat without a second invoice. It is not for anyone shipping a product on top of Claude. It is not for a company with a compliance officer. And it is not for you if you cannot tolerate a bridge that may go dark without notice.

The limitations are real and worth stating plainly:

• One maintainer, young repo. The project's own footprint is small. Bus factor is one until proven otherwise.
• Subscription terms are not yours to reinterpret. Flat-rate consumer plans are not commercial API contracts. Using a Pro/Max login to serve multiple devices and IDEs is exactly the kind of usage that terms-of-service language tends to address. Read Anthropic's terms before you point your whole team at it.
• OpenAI compatibility is a shape, not a guarantee. Tool calling, streaming edge cases, and token accounting will not match OpenAI byte for byte. Expect paper cuts.
• The cache and quotas are conveniences, not governance. Per-key quotas keep your own devices polite. They do not protect you from an upstream rate limit or a policy change.

There is a broader pattern worth naming. The industry keeps building bridges between walled gardens — the same instinct that produced V7 giving AI agents institutional memory and MIT's stranded AIs inventing play to get better at work. Bridges are useful. Bridges are also where the load concentrates when one side moves. OCP is a well-built bridge over a river that changes course.

# Should you run it?

Yes, if you are a hobbyist or a small shop that wants to squeeze more out of a subscription you already pay for, and you are comfortable being your own support desk. Clone it, read the alignment code, issue keys per device, cap the quotas, and keep it on the LAN. No, if you need an SLA, if your usage is commercial in a way your plan does not cover, or if a broken Tuesday would cost you more than the API bill you were avoiding.

The $0 headline is honest about the money and quiet about the risk. That is the trade. OCP is a clever, young, single-maintainer tool that solves a real annoyance — and the annoyance it solves is smaller than the dependency it creates. Know which one you are buying.