Using Your Claude Subscription in Pi Agent — At Your Own Risk

A developer has released pi-sub-anthropic, a Pi extension that lets users point their Claude Pro/Max subscription at the third-party Pi agent harness, circumventing Anthropic's server-side enforcement. The extension reproduces Claude Code's wire fingerprint to present a third-party client as first-party, which risks account termination. The developer advises against using it unless users are prepared to lose their Anthropic account.

If you use Pi https://pi.dev/ and you're willing to gamble your Anthropic account on it, here you go: https://github.com/spksoft/pi-sub-anthropic https://github.com/spksoft/pi-sub-anthropic That sentence is not a joke, and I'd rather lead with it than bury it. I watched Mario Zechner's talk, Building pi in a World of Slop https://www.youtube.com/watch?v=RjfbvDXpFls , and came out the other side genuinely in love with Pi. Four primitives — read, write, edit, bash — a system prompt under a thousand tokens, and everything else is an extension you write yourself. After a couple of years of agents that ship with fifty tools and opinions about all of them, that minimalism lands hard. Then I hit the wall everyone hits. Pi's provider layer does the correct thing: it follows Anthropic's policy and doesn't let you point a Claude Pro/Max subscription at a third-party harness. Since April 4, 2026, Anthropic has enforced this server-side — third-party tools authenticating with a subscription OAuth token get a 400 telling them their usage draws from extra usage, billed per token, not from plan limits. There's an open issue on the Pi repo https://github.com/earendil-works/pi/issues/3372 from someone running into exactly this. That is a reasonable position for Anthropic to hold. Subscriptions were priced for interactive use, not for agents burning tokens unattended, and first-party clients get prompt-cache behaviour that third-party harnesses don't. I'm not writing a grievance post about it. But I was curious. And curiosity, for me, mostly expresses itself as "what happens if I try." So I ported the Anthropic provider layer out of omp oh-my-pi https://github.com/can1357/oh-my-pi , pinned at v17.4.2, into a standalone Pi extension. It registers a second provider — pi-sub-anthropic — alongside Pi's built-in one, so nothing about your existing setup changes until you explicitly select it. pi install npm:pi-sub-anthropic /login pi-sub-anthropic /model pi-sub-anthropic/claude-opus-5 No runtime dependencies, no build step. Pi type-strips the TypeScript at load, and the Messages API is spoken directly over fetch . Nine models registered, mirroring Pi's own Anthropic catalog. It also accepts an API key via PI SUB ANTHROPIC API KEY , which is the boring, sanctioned path — billed as API credits, no fingerprint games, and the one I'd actually recommend to anyone reading this. To make a subscription credential work outside Claude Code, this extension reproduces Claude Code's wire fingerprint: user-agent, beta headers, system-block layout, token clamp, tool naming, billing attestation. That is, by construction, presenting a third-party client as a first-party one . Which means: If you aren't prepared to lose the account, don't install it. Use an API key, or use Pi's built-in anthropic provider, or just use Claude Code — which is, after all, the client this thing is imitating. The interesting work here isn't mine. The OAuth flow, the wire fingerprint, the beta profile, the token clamp, the tool naming, the billing attestation — all of that is omp https://github.com/can1357/oh-my-pi 's discovery. What I contributed is a Node-compatible translation, tests, and documentation. MIT licensed, carrying omp's notice. Not an official Anthropic or Pi project. Use at your own risk.