pi-auto

pi-auto is a Pi package that provides the pi-auto MCP server and a Skill for pi-coding-agent: usage data from @ccusage/pi and tools to switch the active provider by strategy — load-balancing (equalize usage across backends) or high-availability (fixed priority order).

Prerequisites

Install the following (globally) before using pi-auto:

Example: install all globally and set the ccusage alias

# 1. Pi Coding Agent
npm install -g @mariozechner/pi-coding-agent

# 2. Usage tracking (for pi-auto)
npm install -g @ccusage/pi

# 3. Alias for ccusage (pick one: bash or zsh)
echo "alias ccusage='npx @ccusage/pi@latest'" >> ~/.bashrc
# echo "alias ccusage='npx @ccusage/pi@latest'" >> ~/.zshrc

# 4. Pi MCP adapter (run after Pi is installed)
pi install npm:pi-mcp-adapter

# 5. Cursor provider for Pi (optional)
pi install npm:@netandreus/pi-cursor-provider

Then reload your shell (e.g. source ~/.bashrc or source ~/.zshrc) so ccusage is available.

Related MCP server: OpenAI Token Manager MCP

Install (Pi)

Prerequisite: Install the MCP adapter in Pi so it can run MCP servers:

pi install npm:pi-mcp-adapter

Install the pi-auto package (Skill is loaded automatically):

pi install npm:@netandreus/pi-auto

Or from git:

pi install git:github.com/netandreus/pi-auto

MCP config: Add the MCP server to pi’s config so the agent can call the tools.

~/.pi/agent/mcp.json (or your pi MCP config path):

If you installed via pi install npm:@netandreus/pi-auto, pi uses global npm; point to the installed package. Using the binary name (after a global npm install of the package):

{
  "mcpServers": {
      "pi-auto": {
        "command": "pi-auto-mcp",
        "lifecycle": "keep-alive",
        "directTools": true
      }
  }
}

Or with explicit path (replace GLOBAL_NPM_ROOT with the output of npm root -g):

{
  "mcpServers": {
      "pi-auto": {
        "command": "node",
        "args": ["GLOBAL_NPM_ROOT/@netandreus/pi-auto/dist/index.js"],
        "lifecycle": "keep-alive",
        "directTools": true
      }
  }
}

Exposed tools: mcp_pi-auto_pi_get_usage, mcp_pi-auto_pi_suggest_provider, mcp_pi-auto_pi_set_provider, mcp_pi-auto_pi_get_provider, mcp_pi-auto_pi_get_strategy, mcp_pi-auto_pi_set_strategy, mcp_pi-auto_pi_get_priority, mcp_pi-auto_pi_set_priority.

Skill: The Skill is loaded automatically when the package is installed. Invoke with /skill:pi-auto or ask about usage, balancing cost, or switching provider.

MCP in Cursor Agent

To use the pi-auto MCP server from Cursor Agent (not only from Pi), connect the server, enable it, and allow its tools.

1. Connect MCP server to agent

Add the pi-auto MCP server to Cursor’s MCP config (e.g. ~/.cursor/mcp.json):

{
  "mcpServers": {
    "pi-auto": {
      "command": "pi-auto-mcp",
      "lifecycle": "keep-alive",
      "directTools": true
    }
  }
}

2. Enable the MCP server

The server may appear as “not loaded (needs approval)”. Enable and approve it:

agent mcp list
# pi-auto: not loaded (needs approval)

agent mcp enable pi-auto
# ✓ Enabled and approved MCP server: pi-auto

Check that tools are available:

agent mcp list-tools pi-auto
# Tools for pi-auto (8):
# - pi_get_priority ()
# - pi_get_provider (scope, projectPath)
# - pi_get_strategy ()
# - pi_get_usage (period)
# - pi_set_priority (priority)
# - pi_set_provider (provider, model, scope, projectPath)
# - pi_set_strategy (strategy)
# - pi_suggest_provider (period)

3. Allow tools from this MCP

Ensure Cursor Agent is allowed to call pi-auto MCP tools. In ~/.cursor/cli-config.json, under permissions.allow, include:

"permissions": {
  "allow": [
    "Shell(ls)",
    "Mcp(pi-auto:*)"
  ],
  "deny": []
}

Features

Configuration

Requirements: Node 18+, and (for usage data) @ccusage/pi available via npx.

Environment

Server config

Strategy and priority are stored in ~/.pi/agent/pi-auto.json.

Example:

{
  "strategy": "load-balancing",
  "priority": ["codex", "claude-code", "cursorai"],
  "defaultModels": {
    "claude-code": { "provider": "anthropic", "model": "claude-sonnet-4-20250514" },
    "codex": { "provider": "openai", "model": "gpt-4o" },
    "cursorai": { "provider": "cursor", "model": "auto" }
  }
}

Usage strategies

• Load-balancing — pi_suggest_provider picks the backend with the lowest current usage (total tokens). Use pi_get_usage to see per-backend totals; then call pi_suggest_provider and optionally pi_set_provider to switch.

• High-availability — pi_suggest_provider returns the first backend in priority (and its default model). Use pi_set_priority to change the order (e.g. prefer Codex, then Claude, then CursorAI).

When you need to balance cost or ensure a preferred provider, use the pi-auto tools: get/set strategy and priority, get current provider, get usage, suggest a provider, and set it with pi_set_provider.

Backend mapping

Usage from @ccusage/pi (with --json --breakdown) is mapped to three backends:

You can override mapping and default models in ~/.pi/agent/pi-auto.json if your setup differs.

Publishing (npm)

To publish to npmjs.com:

1. Log in: npm login (use an account with access to the @netandreus scope).

2. Bump version in package.json (e.g. 0.1.0 → 0.1.1).

3. Build and publish:
   npm publish
   (runs prepublishOnly → yarn build, then publishes; scoped package is published as public via publishConfig.access.)

To see what will be included: npm pack --dry-run.

License

MIT