pi-claude-code

This Pi extension lets you use your Claude Pro or Max subscription in Pi by routing all Claude model requests through the official Claude Code CLI.

Pi still handles conversation history, approvals, compaction, and tool execution. Claude Code only talks to the model. It never runs tools itself.

Warning

The future of this extension is uncertain. Right now it works by running Pi's requests through the official Claude Code CLI in non-interactive mode (claude -p), and that usage counts against your normal Claude subscription limits. In May 2026 Anthropic announced that claude -p, Agent SDK, and third-party usage would move to a separate, smaller "Agent SDK credit" billed at API rates. Anthropic has since put that change on hold and says it will give advance notice before any future change.

Until then the extension keeps working as described here, but that could change at any time. If Anthropic changes how claude -p usage is billed, this extension may cost more, stop drawing from your subscription, or stop working altogether. See Zed's write-up for more background.

Requirements

• Pi

• Claude Code, installed and on your PATH:

  npm install -g @anthropic-ai/claude-code

• A Claude subscription that you are logged into with Claude Code

Installation

pi install https://github.com/earendil-works/pi-claude-code

Then pick a model from the Claude Code Subscription (Experimental) provider in Pi.

Login and logout

Claude Code handles authentication. Log in and out with the Claude Code CLI:

claude auth login
claude auth logout

The extension uses whatever login Claude Code currently has. You don't need Pi's /login or /logout for this provider. The extension doesn't read, copy, or store your Claude Code credentials.

Run /claude-code-status in Pi to check that Claude Code is installed and logged in.

Configuration

These environment variables are optional:

• PI_CLAUDE_CODE_COMMAND: path or name of the Claude Code executable (default: claude).
• PI_CLAUDE_CODE_CONFIG_DIR: Claude Code config directory to use. The extension passes it to Claude Code as CLAUDE_CONFIG_DIR.
• PI_CLAUDE_CODE_POWERSHELL: PowerShell executable used to run a .ps1 PI_CLAUDE_CODE_COMMAND (default: pwsh when installed, otherwise powershell.exe).

Windows

npm installs Claude Code as claude.cmd (and a claude.ps1 for PowerShell). Both forms work as PI_CLAUDE_CODE_COMMAND:

• .cmd and .bat launchers run through cmd.exe with their arguments escaped once and passed verbatim, so installation paths with spaces and arguments containing quotes or & reach Claude Code unchanged.
• .ps1 launchers run in a noninteractive PowerShell session with -NoProfile and -ExecutionPolicy Bypass. The extension prefers pwsh (PowerShell 7) because Windows PowerShell 5.1 cannot forward every quoted argument from a script to a native program.

Model discovery and privacy

The extension asks Claude Code for the account's model catalog automatically. The discovery probe allows Claude Code's normal feature-flag and catalog traffic, including its usage telemetry, because disabling feature flags hides models such as Mythos and previous-generation models. This is a known Claude Code issue. The probe sends no conversation history or generation request. Automatic CLI updates, official marketplace auto-install, feedback uploads, and error reporting remain disabled.

To opt out of discovery telemetry, set DISABLE_TELEMETRY=1 before starting Pi. Discovery also preserves inherited CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC, DO_NOT_TRACK, and DISABLE_GROWTHBOOK settings. These can reduce the discovered catalog; the extension does not remove them to restore missing models. Claude Code treats any non-empty DISABLE_TELEMETRY or CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC value as enabled, including 0 or false; unset them to allow normal discovery again. See Claude Code's environment variables.

Conversation workers still disable telemetry and nonessential traffic regardless of the discovery settings.

Usage and billing

Requests made through this extension count against your Claude subscription's usage quota and rate limits, the same as any other Claude Code usage. The extension does not give you separate or unlimited usage.

The extension ignores ANTHROPIC_API_KEY in Claude Code child processes without changing your environment. To avoid silently changing how you're authenticated or billed, it still rejects auth-token, proxy, Bedrock, Vertex, or Foundry overrides for Claude Code.

Limitations

• The extension is experimental.
• Branching, editing earlier messages, or changing settings mid-conversation starts a fresh Claude Code process that replays the conversation. The first response after that can be slower and may use more quota.
• Complex tool schemas (nullable types, oneOf, allOf, anyOf) may be simplified before Claude sees them.
• On Windows, Windows PowerShell 5.1 mangles quoted arguments (such as the inline MCP configuration JSON) when a .ps1 launcher forwards them to a native program. Install PowerShell 7 or use the claude.cmd launcher in that case.
• Adding or removing tools in the middle of a conversation keeps the prompt cache only on supported models. On other models it may use more quota.

For details on how the extension works internally, see ARCHITECTURE.md.