🧳 pi-mcporter

Use MCP tools from pi through one stable tool (mcporter), powered by MCPorter.

🧠 Philosophy: CLI > MCP

• pi is fundamentally CLI-first.
• Prefer native CLIs whenever available (gh, git, kubectl, aws, etc.).
• Use MCP via mcporter when it adds clear value (for example: Linear, Slack, hosted auth-heavy integrations, cross-tool workflows).
• This package is intentionally a thin bridge, not a replacement for pi’s normal tool-driven workflow.

✨ Why use this package

• Keeps context small: one stable mcporter tool instead of exposing many MCP tools
• Uses your MCPorter config/runtime as source of truth
• Supports discovery (search), schema help (describe), and execution (call)
• Returns useful error hints for auth/offline/http/stdio failures

📋 Prerequisites

You need MCPorter installed and configured with at least one MCP server:

npm install -g mcporter
npx mcporter list          # verify your servers are visible

📦 Install

Install as a pi package:

pi install npm:pi-mcporter

Try it once without installing:

pi -e npm:pi-mcporter

🚀 Quick start

1. Confirm MCPorter sees your servers:

npx mcporter list

2. Start pi:

pi

3. Ask for what you need — pi picks the right MCP tools automatically:

• What are my open Linear issues this sprint?
• Catch me up on #engineering in Slack from today.
• Find the onboarding runbook in Notion and summarize the setup steps.

🔍 The three actions

The mcporter tool has three actions that map to a natural discovery → execution workflow.

search — find tools by keyword

Use when you don't know the exact server or tool name.

{ "action": "search", "query": "linear issue", "limit": 5 }

Returns matching selectors with short descriptions:

linear.create_issue — Create a new issue in a Linear team
linear.list_issues  — List issues matching a filter

describe — get the full schema for a tool

Use when you know the selector but need to see its required parameters before calling.

{ "action": "describe", "selector": "linear.create_issue" }

Returns the full JSON Schema for the tool's input, including required vs. optional fields and their types.

call — invoke a tool

Use once you know the selector and its schema.

{
  "action": "call",
  "selector": "linear.create_issue",
  "args": { "title": "Fix login bug", "teamId": "TEAM-1", "priority": 2 }
}

For arguments that are awkward to express as nested JSON, you can pass them as a JSON string via argsJson instead of args.

Typical workflow

search "linear issue"          →  discover: linear.create_issue
describe linear.create_issue   →  learn required fields: title, teamId
call linear.create_issue       →  execute with those fields

In practice pi follows this pattern automatically. With mode: "preload" the catalog is already warm at agent start, so pi can often skip search/describe and jump straight to call.

🧰 Tool input (reference)

Tool name: mcporter

• action: "search" | "describe" | "call"
• selector?: "server.tool" (required for describe and call)
• query?: free-text query for search
• limit?: result limit (default 20, max 100)
• args?: object arguments for call
• argsJson?: JSON-object-string fallback for call
• timeoutMs?: per-call timeout override

⚙️ Configuration

Configure the extension in ~/.pi/agent/mcporter.json:

{
  "configPath": "/absolute/path/to/mcporter.json",
  "timeoutMs": 30000,
  "mode": "lazy"
}

• MCPORTER_CONFIG=/absolute/path/to/mcporter.json still overrides configPath from the settings file.
• configPath: optional explicit MCPorter config path. If omitted, MCPorter uses its normal default resolution.

... View full README on GitHub