ToS check: Is running Claude Code agents through Paperclip within Anthropic's terms?

[matthewpomar]

Hey! Really digging what you're building here!

Quick question before I dive in: where does Paperclip stand with Anthropic's terms of service?

Earlier this year Anthropic cracked down hard on third-party tools that were using subscription OAuth tokens to access Claude outside the official clients. They banned accounts, updated their legal docs, and even sent legal requests to projects like OpenCode to remove Claude subscription support entirely.

From what I can see, Paperclip does this the right way. It launches the actual Claude Code binary through adapters rather than intercepting tokens or spoofing the client. So the auth layer looks clean.

The part I'm less sure about is usage patterns. Anthropic's docs say Pro and Max limits assume "ordinary, individual usage." Running a fleet of Claude Code agents autonomously on heartbeat schedules is... not that. Has anyone gotten clarity from Anthropic on whether orchestrated, unattended Claude Code sessions are cool, or whether that falls under the automated access restrictions in the Consumer ToS (further clarified in the Claude Code legal compliance docs).

One related question: right now it looks like the Claude Code adapter only supports terminal login (subscription OAuth), not API key auth. Are there plans to support ANTHROPIC_API_KEY? That would let users opt into usage-based billing and Commercial Terms protections, which seems like the cleaner path for running agents autonomously at scale.

Not trying to create drama. I think the adapter approach is solid. Just want to make sure I'm not going to wake up to a banned account after letting my agents run overnight.

Thanks for the awesome project! 🙌

[Augustas11]

Great question and you've clearly done your homework on the adapter mechanics — the distinction you're drawing is real and important.

On the adapter approach

You're correct that the claude_local adapter launches the actual Claude Code binary rather than intercepting OAuth tokens or spoofing the client. This is meaningfully different from tools that were targeted in earlier enforcement actions. Paperclip doesn't touch the auth layer — the agent authenticates directly with Anthropic the same way a human would in a terminal session.

On the "ordinary individual usage" question

This is the honest answer: we don't have explicit written confirmation from Anthropic that autonomous, unattended, heartbeat-scheduled Claude Code sessions are within the Consumer ToS. The ToS language around "ordinary individual usage" is ambiguous for this use case. Our read is that the intent of those restrictions is to prevent token-sharing across accounts and API arbitrage — not to prohibit a single user running their own agents autonomously. But we can't represent that definitively.

The safest path for production fleet use is exactly what you suggested: ANTHROPIC_API_KEY mode. This puts you on the Commercial Terms rather than the Consumer ToS, which are explicitly designed for programmatic, unattended use. It also gives you cleaner cost attribution through the API rather than subscription.

On ANTHROPIC_API_KEY support in the Claude adapter

The claude_local adapter currently relies on the Claude Code CLI's own auth (terminal login / subscription OAuth). Direct API key injection is tracked — if you want to follow progress or add pressure, see issue #37 and the related PR work around LiteLLM gateway support (#1191) which provides an alternative path for API-key-based routing today.

If running autonomous agents at scale on a subscription account makes you nervous, the pragmatic interim path is: use Claude Code's --model flag pointed at an API-backed provider via LiteLLM (PR #1191), or wait for native ANTHROPIC_API_KEY support in claude_local.

We'll add a note to the docs clarifying the ToS landscape and the recommended production path. Appreciate you raising this properly rather than just hoping for the best.