Does Anthropic's Agent SDK Change Affect Callipso?
On June 15, 2026, Anthropic moves Agent SDK and claude -p usage to a separate credit. Here is why it does not affect Callipso or your subscription limits.
Does Anthropic's Agent SDK Change Affect Callipso?
On June 15, 2026, Anthropic changes how paid Claude plans handle programmatic usage. The short version: the Claude Agent SDK and the claude -p CLI command no longer draw from your subscription's interactive limits — they draw from a separate, smaller monthly credit. Tools built on the Agent SDK are affected. Callipso is not. This post explains exactly why, because the distinction is precise and worth getting right.
What Anthropic Announced
Starting June 15, 2026, programmatic usage of Claude gets its own budget instead of sharing your subscription rate limits. According to Anthropic's help center, this covers the Claude Agent SDK, the claude -p CLI command, Claude Code GitHub Actions, and third-party apps built on the Agent SDK. Eligible subscribers receive an email on June 8 and must manually claim the credit. It resets each billing cycle and does not roll over.
| Plan | Monthly Agent SDK credit | When exhausted |
|---|---|---|
| Pro ($20) | $20 | Pay API rates via usage credits, or programmatic usage pauses |
| Max 5x ($100) | $100 | Same |
| Max 20x ($200) | $200 | Same |
Interactive use inside Anthropic's own interfaces — the claude terminal UI, the desktop app, claude.ai — is explicitly unchanged. Your subscription limits stay exactly as they are for that.
Programmatic vs Interactive: Where the Line Is
The useful question is not "is there a tool layered on top of Claude Code." It is "what makes the request to Anthropic's servers." That is the line Anthropic actually drew.
When you run claude -p or call the Agent SDK, your code is the client. Your script, or the app built on the SDK, is sitting directly in the inference request path — it constructs the call, it talks to Anthropic, it gets the response. That is what "programmatic" means in this announcement, and that is what moved to the separate credit.
When you type into the genuine claude terminal UI, Anthropic's own unmodified client makes the request. That is "interactive." The distinction is about the request path, not about whether software is involved at all.
Why Callipso Is Not Affected
Callipso is never in the inference request path. It does not import the Agent SDK, and it does not invoke claude -p anywhere. Callipso is a transparent orchestrator: it discovers the genuine claude sessions you launch yourself across your terminals and IDEs, then automates the input side and observes the output side.
Concretely, Callipso drives the genuine interactive claude UI by delivering text into it — the same UI you would type into by hand — and it reads Claude Code's hook events (prompt-start, stop, PostToolUse, PermissionRequest) to track session state. The request to Anthropic still originates from Anthropic's own unmodified client: genuine system prompt, genuine harness, genuine caching. Callipso sits between you and the keyboard, not between Claude Code and Anthropic.
This is a categorical difference from SDK-wrapper tools, not a difference of degree:
| Tool | How it reaches Anthropic | June 15 impact |
|---|---|---|
Agent SDK apps, claude -p scripts, CI integrations | Their code is the client — in the request path | Draws from the separate Agent SDK credit |
| ACP adapters and SDK-based UI wrappers | Built on the Agent SDK — in the request path | Draws from the separate Agent SDK credit |
| Callipso | Drives the genuine claude UI via text delivery; reads hooks | None — this is interactive use |
There is also no authentication surface to worry about. Callipso never handles your Anthropic credentials. Every Callipso user runs their own claude install, authenticated by themselves, with their own subscription. Callipso does not see, store, broker, or resell access to anyone. There is nothing for the June 15 change to detect or reauthorize.
FAQ
Will Callipso users lose subscription limits on June 15, 2026?
No. The change applies to the Claude Agent SDK and the claude -p CLI command — programmatic paths where your own code is the client talking to Anthropic. Callipso drives the genuine interactive claude UI and reads hook events; it is never in the inference request path. Callipso usage is interactive use, which Anthropic explicitly left unchanged.
Does Callipso use the Claude Agent SDK?
No. The Agent SDK is not a dependency of Callipso, and Callipso does not invoke claude -p. Cloning a session launches the genuine interactive claude UI with the first prompt pre-filled — equivalent to running claude "your prompt" yourself in a terminal, which opens the real UI rather than a headless call.
How is Callipso different from T3 Code, Zed's ACP adapter, or OpenClaw?
Those tools are built on the Agent SDK or claude -p — their code makes the request to Anthropic, so they fall under the new programmatic credit. Callipso does not put itself in the request path. It orchestrates the genuine Claude Code UI from the outside via text delivery and hooks, which keeps it on the interactive side of the line.
Do I need to do anything before June 15?
Nothing for Callipso to keep working. If you separately use the Agent SDK or claude -p for your own scripts or CI, watch for Anthropic's June 8 email and claim that credit. That is independent of Callipso.
What We Are Watching
We will be honest about where the uncertainty is. The technical picture is clear: Callipso has no programmatic path into Anthropic, so the June 15 mechanism has nothing to act on. The grayer question is policy intent. Anthropic's framing emphasizes interactive, human-in-the-loop use of its own interfaces, and a multi-session orchestrator is not the narrowest reading of that.
Three things keep that a question of vibe rather than a rule. The announcement, as written, names the Agent SDK and claude -p specifically and nothing else. There is no mechanism to enforce a broader reading against Callipso, because Callipso's traffic is the genuine interactive client. And any rule broad enough to catch Callipso would also catch terminal multiplexers and keyboard-macro tools — which Anthropic did not address. We will track Anthropic's clarifications and update this post if the picture changes.
There is a deeper point here. Callipso sits outside the request path not by luck but by design. Discovering agents across the terminals and IDEs you already use cannot be built on the Agent SDK in the first place — the SDK only sees processes it spawned itself. The architecture that makes the orchestration category work is the same architecture that keeps Callipso clear of this change.
Closing
Anthropic's June 15 change moves programmatic Claude usage — the Agent SDK and claude -p — into a separate credit. Callipso is interactive use: it drives the genuine claude UI and reads hooks, and it is never the client making the request to Anthropic. Your subscription limits work with Callipso on June 15 exactly as they do today. If you want to try it, the download is here.