pi-anyrouter

A pi provider extension that adapts requests to the client-specific shapes accepted by AnyRouter.

It currently supports both Claude Code / Claude Agent SDK requests and Codex Responses Lite requests. It does not require a separate local relay process.

Status

Confirmed working with provider anyrouter using Claude and Codex model routes:

pi --model anyrouter/claude-opus-4-8 --no-session --no-tools -p "Reply with exactly OK"
pi --model anyrouter/gpt-5.6-sol --no-session --no-tools -p "Reply with exactly OK"

Expected output:

OK

What this package fixes

Some AnyRouter Claude endpoints, especially claude-opus-4-6, reject older or more generic Claude-style requests with errors like:

{"error":{"type":"new_api_error","message":"invalid claude code request (...)"},"type":"error"}

This package selects an adapter by model family:

• Claude models use POST /v1/messages?beta=true with Claude Code headers, system blocks, metadata, and Anthropic SSE conversion.
• Codex models use POST /v1/responses with the Codex Responses Lite headers/body shape and OpenAI Responses SSE conversion.

Features

• Registers provider: anyrouter
• Supports Claude Code and Codex Responses Lite request validation
• No local relay/proxy required
• Converts tool names to Claude Code naming
• Supports reasoning via thinking + output_config
• Built-in request/response debug dump
• Retries transient upstream failures like HTTP 520/502/503/504 automatically
• Packaged so it can be shared as a pi package

Install

Option A: local development checkout

Clone into a normal directory:

git clone https://github.com/xifan2333/pi-anyrouter.git
cd pi-anyrouter

Install as a pi package from the local path:

pi install .

Or install directly from a path later:

pi install /absolute/path/to/pi-anyrouter

Option B: direct git install

Once pushed to GitHub, install with:

pi install git:github.com/xifan2333/pi-anyrouter

Or pin a ref/tag:

pi install git:github.com/xifan2333/pi-anyrouter@<tag>

Option C: manual extension placement

If you only want the extension file layout, place it at either:

• ~/.pi/agent/extensions/anyrouter/index.ts
• .pi/extensions/anyrouter/index.ts

Then reload pi:

/reload

Config

Create:

• ~/.pi/agent/anyrouter.json

Example:

{
  "baseUrl": "https://anyrouter.top",
  "apiKey": "YOUR_ANYROUTER_KEY_OR_ENV_NAME",
  "models": [
    {
      "id": "claude-opus-4-8",
      "name": "Claude Opus 4.8",
      "reasoning": true,
      "input": ["text"],
      "cost": {
        "input": 0,
        "output": 0,
        "cacheRead": 0,
        "cacheWrite": 0
      },
      "contextWindow": 200000,
      "maxTokens": 32000
    }
  ]
}

You can also override config values with environment variables:

• PI_ANYROUTER_CC_CONFIG
• PI_ANYROUTER_CC_BASE_URL
• PI_ANYROUTER_CC_API_KEY
• PI_ANYROUTER_CC_MAX_RETRIES (default: 10)
• PI_ANYROUTER_CC_STREAM_MODE (default: force)

Notes:

• apiKey can be a literal key or an environment variable name.
• shell-command values like "!command" are intentionally not supported.
• this package reads ~/.pi/agent/anyrouter.json by default rather than models.json.
• PI_ANYROUTER_CC_STREAM_MODE=auto tries real SSE streaming first and falls back to the old non-stream JSON path if streaming fails before any content arrives.
• PI_ANYROUTER_CC_STREAM_MODE=force uses SSE only; PI_ANYROUTER_CC_STREAM_MODE=off disables SSE.

Use

After installation/configuration:

/reload
/model

Choose a configured model, for example:

• anyrouter / claude-opus-4-8
• anyrouter / gpt-5.6-sol

Or run directly:

pi --model anyrouter/gpt-5.6-sol

Debugging

Enable request/response dumps:

PI_ANYROUTER_CC_DEBUG=1 pi --model anyrouter/gpt-5.6-sol -p "Reply with exactly OK"

Optional custom debug directory:

PI_ANYROUTER_CC_DEBUG=1 \
PI_ANYROUTER_CC_DEBUG_DIR=/tmp/anyrouter-cc-debug \
pi --model anyrouter/gpt-5.6-sol -p "Reply with exactly OK"

Default debug output directory:

• .pi/anyrouter-cc-debug/

If AnyRouter intermittently returns HTTP 520 / Origin Error, this package retries transient upstream failures automatically with exponential backoff. You can tune the retry count with PI_ANYROUTER_CC_MAX_RETRIES.

The runtime uses SSE by default (PI_ANYROUTER_CC_STREAM_MODE=force). For troubleshooting, you can force old behavior with:

PI_ANYROUTER_CC_STREAM_MODE=off pi --model anyrouter/claude-opus-4-8 -p "Reply with exactly OK"

Share as a pi package

This repo is already structured as a pi package through package.json:

{
  "keywords": ["pi-package"],
  "pi": {
    "extensions": ["./index.ts"]
  }
}

That means users can install it with:

pi install git:github.com/xifan2333/pi-anyrouter

Publish checklist

Publish to GitHub

git init
git add .
git commit -m "feat: add AnyRouter Claude and Codex adapters"
git branch -M main
git remote add origin [email protected]:xifan2333/pi-anyrouter.git
git push -u origin main

Optional npm publish

If you want npm distribution too:

npm publish

Then users can install with:

pi install npm:pi-anyrouter

Important behavior

This package implements the Claude Code / Claude Agent SDK and Codex Responses Lite request shapes currently accepted by AnyRouter.

If AnyRouter changes its validation again, enable debug dumps and compare the request shape against fresh official client captures.

Credits

Based on pi-anyrouter-cc by zhenliangzhang, licensed under the MIT License.

This is an unofficial community project and is not affiliated with Anthropic, OpenAI, or AnyRouter.

License

MIT. See LICENSE for the original and modification copyright notices.