pi-claude-subscription

Use your Claude Pro/Max subscription in pi. This pi provider runs Claude models through the Claude Agent SDK, while pi keeps its own agent loop, tools, system prompt, compaction and session tree. Claude Code serves only as the authenticated model transport.

pi install npm:pi-claude-subscription
# or: pi install git:github.com/jkeskikangas/pi-claude-subscription
/model claude-sdk/claude-sonnet-5-5

On the bundled benchmark it matches Claude Code's task success, processes 5–10× fewer input tokens and costs about half as much (API-equivalent). See bench/RESULTS.md.

Requirements: pi ≥ 1.0 and a logged-in Claude Code (claude → /login). The extension uses Claude Code's own login. It never reads pi's credentials. It also removes ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN and the Bedrock/Vertex/Foundry switches from Claude Code's environment, so none of them can move you to API billing.

How it works

sequenceDiagram
    participant L as pi agent loop
    participant P as provider
    participant C as Claude Code process<br/>(one per conversation)
    Note over C: system prompt = pi's prompt<br/>tools = pi's tools via in-process MCP
    L->>P: streamSimple
    P->>C: streaming input
    C-->>P: raw API stream events
    P-->>L: toolCall events
    Note over C: MCP call parked
    L->>L: pi runs the tool
    L->>P: streamSimple (tool result)
    P->>C: tool result releases the parked MCP call
    C-->>P: stream continues

• pi's prompt and tools, not Claude Code's. The Claude Code process runs with systemPrompt set to pi's prompt, tools: [], and pi's tools as an in-process MCP server. A request carries pi's prompt plus a two-line Claude Code preamble, not Claude Code's ~30k-token prompt and tool set.
• One live process per conversation. pi's next provider call after a tool round delivers the tool results to the parked MCP handlers, matched by tool_use id, and Claude Code continues the same API conversation. New user turns go in through streaming input. Nothing is respawned and the prompt cache stays warm.
• Steering is native. A message you type while a tool runs is queued in Claude Code before the tool results are released, so the model sees it at that tool boundary as Claude Code's own "user sent a message while you were working" notice.
• Divergence forks Claude Code's own transcript. pi's history can stop matching the live process after an abort, /tree navigation, a fork, a thinking-level change or a pi restart. The provider then looks up the longest matching prefix in an anchor registry (pi transcript hash → Claude Code session and message), using the transcript-entry uuid streamed with that assistant message, and calls the Agent SDK's forkSession to copy the transcript up to that entry before resuming the fork. Claude Code replays its own bytes, so the prompt cache still hits. The old process exits before the fork reads its transcript. Anchors persisted by older versions have no streamed uuid; their entry uuid is resolved from Claude Code's transcript file as a fallback.
• Foreign history is synthesized. Turns from other providers, a compacted session, or a session from before the extension was installed are written as a Claude Code transcript (tool ids sanitized, cross-model thinking dropped) and resumed.
• One-shot requests, such as compaction summaries and branch summaries, run tool-less and unpersisted.
• pi owns context. Claude Code's auto-compaction, CLAUDE.md loading, auto-memory, title generation and other non-essential traffic are disabled. The MCP output cap and tool timeout are lifted so pi's own truncation and timeouts apply.

System prompt

For claude-sdk models the extension replaces pi's default prompt prefix with one written for Claude. It lists the selected tools and keeps every tool and prompt guideline that pi and other extensions contribute, adds Claude-oriented working rules (read before editing, parallel tool calls, verify, report faithfully) and points to pi's docs. pi's context files, skills and working directory are unchanged, and your own SYSTEM.md always wins.

Set "promptMode": "pi" in the global claude-sdk.json, or PI_CLAUDE_SDK_PROMPT=pi, to keep pi's default prompt.

Billing. Requests declare themselves honestly as Agent SDK traffic (sdk-ts). An inherited Claude Code entrypoint is always cleared. Observed on a Max plan on 2026-10-04: with the Claude prompt, requests ran within the plan's usage limits; with pi's default prompt, Anthropic answered 400 Third-party apps now draw from your extra usage. Anthropic's routing can change; see their Agent SDK plan article.

Configuration

Optional claude-sdk.json in pi's agent directory (~/.pi/agent, or $PI_CODING_AGENT_DIR):

{ "pathToClaudeCodeExecutable": "/opt/homebrew/bin/claude", "maxSessions": 3, "idleMinutes": 20 }

A project's .pi/claude-sdk.json can set maxSessions and idleMinutes only. pi reads that file before you grant project trust, so a project cannot choose which program runs. PI_CLAUDE_SDK_CLAUDE_PATH overrides the executable path.

• PI_CLAUDE_SDK_DEBUG=/path/log: routing decisions (live / resume / synth / fresh).
• Thinking levels map to Claude effort (low … max) on adaptive-thinking models and to thinking budgets on older ones.
• Claude Code writes one-hour prompt-cache entries, so pi's cache warmer is not used. cacheRetention: "none" disables caching.

Tests

• npm test: offline tests. They cover the helpers, plus event translation and routing against a scripted fake Claude Code session: tool loop, Claude Code's own continuations, mid-stream retry, abort, errors, live/resume/synth routing and the busy guard.
• npm run test:live: the provider against real Claude Code: tool loop, multi-turn, restart resume, abort recovery, foreign history, tool-less requests. With PROXY=http://127.0.0.1:8787 it also asserts cache hits.
• npm run test:e2e: pi's own agent loop over RPC: steering, /compact, thinking switch, image input.

Benchmark

bench/proxy.mjs is a logging reverse proxy for api.anthropic.com. Both harnesses run with ANTHROPIC_BASE_URL pointing at it, so every request's real usage, side calls included, is recorded per run.

• bench/run.mjs: 8 coding tasks on a fixture repo with hidden verification tests.
• bench/session.mjs: one 5-turn interactive session per harness.
• bench/report.mjs: summary tables.

Results are in bench/RESULTS.md.

Limitations

• Tool names reach the model as mcp__pi__<name>. pi's tool descriptions are unchanged.
• Claude Code adds a short preamble (billing header, SDK identity), the account email and an environment block. These are about 250 stable tokens.
• Using a subscription through the Agent SDK is subject to Anthropic's terms for your plan.