ptools

ptools turns many upstream MCP servers into one Code Mode tool surface. The model sees a small set of tools for discovery and execution, while host-side ptools keeps the original MCP registry, tool schemas, and dispatch path authoritative.

There are two installable alpha paths:

• @ptools/cli for MCP hosts such as Claude Code and OpenCode
• @ptools/host-node plus @ptools/agent-tools for app code using the AI SDK

MCP Server

Use the MCP server when you want a host to load one ptools server that proxies multiple upstream MCP providers.

npx -y @ptools/cli mcp serve --host node --config .ptools/config.json

Example .ptools/config.json:

{
  "mcpServers": {
    "exa": {
      "url": "https://mcp.exa.ai/mcp"
    },
    "notion": {
      "url": "https://mcp.notion.com/mcp"
    },
    "sheets": {
      "url": "https://mcp.gumloop.com/gsheets/mcp"
    }
  }
}

Claude Code project config:

{
  "mcpServers": {
    "ptools": {
      "type": "stdio",
      "command": "npx",
      "args": [
        "-y",
        "@ptools/cli",
        "mcp",
        "serve",
        "--host",
        "node",
        "--config",
        ".ptools/config.json"
      ]
    }
  }
}

OpenCode config:

{
  "mcp": {
    "ptools": {
      "type": "local",
      "command": [
        "npx",
        "-y",
        "@ptools/cli",
        "mcp",
        "serve",
        "--host",
        "node",
        "--config",
        ".ptools/config.json"
      ],
      "enabled": true,
      "timeout": 30000
    }
  }
}

The host sees ptools' stable Code Mode tools: search_providers, search, get_tool_schema, and execute, plus auth helpers. Upstream provider tools are discovered through search and called from generated JavaScript passed to execute.

AI SDK

Use the AI SDK package when you want to embed ptools in your own app:

npm install @ptools/agent-tools @ptools/host-node

import { generateText, stepCountIs } from "ai";
import { openai } from "@ai-sdk/openai";
import { createNodeCodeModeClientFromConfigFile } from "@ptools/host-node";
import { makePtoolsSession } from "@ptools/agent-tools";
import { toAISDKTools } from "@ptools/agent-tools/ai-sdk";

const ptools = makePtoolsSession(
  await createNodeCodeModeClientFromConfigFile(),
);

try {
  const result = await generateText({
    model: openai("gpt-5.4"),
    tools: toAISDKTools(ptools),
    stopWhen: stepCountIs(8),
    prompt: "Use the configured MCP servers to answer this.",
  });

  console.log(result.text);
} finally {
  await ptools.close();
}

Package docs:

• @ptools/cli: command-line composition surface for MCP hosts
• @ptools/mcp-server: host-neutral MCP stdio adapter
• @ptools/agent-tools: user-facing AI SDK session and adapter package
• @ptools/config: shared config parsing, validation, resolution, and hashing
• @ptools/host-node: Node host layers and Code Mode client factories
• @ptools/mcp-registry: upstream MCP connection, discovery, and dispatch
• @ptools/code-mode: Code Mode search, schema, and execute orchestration
• @ptools/executor: local JavaScript execution host

See packages/mcp-server/README.md and packages/agent-tools/README.md for the alpha user guides.