@mpilot/tools

Framework-agnostic Concierge tool registry. The single source of truth for the ConciergeTool interface — every adapter (Vercel AI SDK, OpenAI, LangChain, Coinbase AgentKit, MCP server, React UI) consumes the same shape, with one inputSchema + outputSchema definition feeding all surfaces.

Quickstart

import { createConciergeTools, tool } from '@mpilot/tools';
import { z } from 'zod';

// 1. Define a tool — `tool()` is an identity helper that preserves type inference.
const supply = tool({
  name: 'supply',
  description: 'Supply USDC to Aave V3',
  inputSchema: z.object({ amount: z.number() }),
  outputSchema: z.object({ txHash: z.string() }),
  invoke: async ({ amount }) => ({ txHash: `0x${amount.toString(16)}` }),
  supportsNetwork: (id) => id === 5000, // Mantle Mainnet only
});

// 2. Provider packages export a `tools(agent)` factory that returns ConciergeTool[].
const aaveTools = () => [supply];

// 3. Aggregate. Network-incompatible tools are filtered out automatically.
const tools = createConciergeTools({ chainId: 5000 }, [aaveTools]);

Adapter pattern

Each adapter package wraps ConciergeTool[] into the framework's native shape. See @mpilot/vercel-ai, @mpilot/openai, @mpilot/langchain, @mpilot/agentkit, @mpilot/mcp for ~30-LOC reference implementations.

Card schemas

/serializable exports the four canonical card schemas the MCP server + React UI parse-then-render against (per ADR-017):

import {
  SerializableProposalCardSchema,
  SerializableTickCardSchema,
  SerializablePortfolioCardSchema,
  SerializableReputationCardSchema,
} from '@mpilot/tools/serializable';

A tool whose job is to emit one of these cards SHOULD use the schema as its outputSchema so the MCP structuredContent round-trip is type-safe.