MCP 服务器

One Agent Plugin

Provides a remote MCP server with four tools to list, search, and execute actions across 500+ integrated apps, using OAuth and platform-specific knowledge packages to guide agents.

README

One, as an Agent Plugin

500+ apps in your agent, packaged to the Agent Plugins 1.0.0 standard. One directory that ChatGPT, Codex, Cursor, GitHub Copilot, Kiro and VS Code all read the same way.

one/
├── plugin.json      # manifest
├── mcp.json         # One's remote MCP server, OAuth, nothing to install
└── skills/
    ├── one-integrations/    # how to act in real apps without guessing parameters
    └── integration-code/    # how to write integration code against real schemas

What you get

The MCP server exposes four tools, no matter how many apps you connect:

Tool	What it does
`list_one_integrations`	Lists platforms and connections, each with the access it allows
`search_one_platform_actions`	Searches actions on a platform
`get_one_action_knowledge`	Returns an action's real parameters, types, and auth
`execute_one_action`	Runs the action against the live account

The two skills are the half that a bare MCP entry does not give you. They teach the agent the search, read, then execute discipline that keeps it from inventing a field name and writing the wrong thing to a real account, and they tell it what to do when a platform is not connected or a call comes back 401.

Auth

There is no key in this package. The MCP server uses OAuth: the first connect opens a browser, and the consent screen is where you scope the agent to specific connections, read-only, or knowledge-only mode. The Agent Plugins spec forbids secrets in headers and env (§7.2.1, §9.2), which rules out shipping a token here even if we wanted to.

One plugin per platform

platforms/ holds a conformant plugin for every platform in One's catalog, generated from the knowledge base:

platforms/one-quickbooks/
├── plugin.json
├── mcp.json
└── skills/quickbooks/SKILL.md

Each skill carries that platform's real action ids, HTTP methods and paths, so an agent can go straight to reading an action's parameters instead of spending a turn on search, and cannot invent an action that does not exist. The tables are generated, never hand-written.

Regenerate them:

ONE_SECRET=... npm run generate            # every platform
ONE_SECRET=... npm run generate -- --platform shopify

Every generated plugin is validated before it is written; anything that fails is dropped and reported rather than shipped. platforms/index.json lists what was produced.

These plugins are derived from One's knowledge base, so they carry the One-Knowledge-1.0 license and its attribution requirement rather than MIT. The root plugin and the tooling are MIT.

Verify conformance

npm install
npm run validate                    # this plugin
npm run validate platforms/one-*    # every generated plugin

The validator checks plugin.json and mcp.json against the official 1.0.0 JSON Schemas (vendored under schemas/, since the spec forbids fetching a schema at load time), then checks the requirements the schemas cannot express: name constraints (§5.5), plugin root containment (§4.1), HTTPS and no-credentials rules for remote servers (§7.2.1), reserved environment names (§9.2), spec version match between the two files (§10.1), and Agent Skills frontmatter on every skill.