@pi-plugins/claude-cli

Use a Claude subscription in pi through the official Claude Code CLI. Every request is made by Claude Code itself and draws from your plan's usage limits, like Claude Code's own requests.

[!NOTE] Anthropic decides how requests from claude --print are billed. In May 2026 it announced that they and third-party apps would move to a separate credit, and in June it paused that change. They still draw from your plan's limits for now. If Anthropic changes this, the plugin may cost more or stop working, though it stops as soon as a request is charged to usage credits unless you allow them.

How it works

pi's built-in Anthropic provider builds each request and reads the response, just as it does for the anthropic provider. Only sending the request is left to Claude Code. Every request runs in its own claude --print process, which signs the request and sends it with your subscription, so it counts against your plan's limits like Claude Code's own requests. The process is stopped once the response is complete and Claude Code has reported which limits paid for it, so Claude Code never runs a tool or sends a request of its own.

Claude Code only signs you in, signs the request and sends it. Everything else stays with pi:

• Minimal system prompt. The model gets pi's system prompt unchanged, after only Claude Code's billing header and a one-line identity. Claude Code's own prompt, project instructions, skills and environment details are left out.
• pi's tools. The model sees pi's tools under their own names, and pi runs them. Claude Code's built-in tools and MCP servers stay off.
• pi's session. pi keeps the conversation, compaction, retries and thinking level. The requests of a pi session share one Claude Code session ID, but Claude Code stores nothing.
• pi's prompt caching. The conversation, tools, thinking settings and cache breakpoints are those of pi's Anthropic provider, so each request reads the earlier conversation from cache. On models that support it, a change to the system prompt or the tools during a session is sent as a message, which keeps the cached conversation valid.
• No copied request details. The plugin pins no headers, version numbers or signing values from Claude Code. Claude Code builds and signs every request itself.

Requirements

The claude CLI must be on your PATH and signed in to a Pro or Max plan:

claude auth login

Tested with pi 1.2.0 and Claude Code 2.1.299.

Install

pi install npm:@pi-plugins/claude-cli

To try it for one session without changing your settings:

pi -e npm:@pi-plugins/claude-cli

Usage

Pick a model from the claude-cli provider with /model, or start pi with one:

pi --model claude-cli/claude-sonnet-5-5

The provider offers the same Claude models as pi's built-in anthropic provider.

Usage credits

When Claude Code reports that a response was charged to usage credits instead of your plan's limits, pi stops with an error rather than sending more requests. This happens once a plan limit is used up while usage credits are turned on, and on every request with models or modes that always use credits. pi also stops without retrying when Anthropic refuses to bill a request to your plan's limits.

To stay on your plan's limits, turn off usage credits at claude.ai/settings/usage. To let pi use them, set PI_CLAUDE_CLI_ALLOW_USAGE_CREDITS=1 before starting pi.

Notes

• Requests are billed to the account claude is signed in to, not to pi's /login. The usage plugin reads the limits of claude-cli models from that account too, so pi needs no Claude sign-in of its own.
• API keys in the environment, such as ANTHROPIC_API_KEY, are not passed on to Claude Code, so requests always use the subscription.
• Cache retention follows pi's PI_CACHE_RETENTION setting. It is five minutes by default and one hour with long.
• When another Claude Code process is refreshing the login, a request waits and tries again for up to a minute.
• Starting Claude Code takes about a quarter of a second, so from its second request on, each session keeps one process started ahead for its next request. An idle process takes about 200 MiB. It is stopped after five idle minutes and when the session ends.