One

One MCP Server

Website · Docs · Dashboard · Changelog · X · LinkedIn

Connect your AI agents to 1000+ apps through a single MCP server. Find the actions a task needs with their documentation, and execute API calls across platforms, without having to manage OAuth tokens or API keys.

npm install -g @withone/cli

one init

That's it. The One CLI will prompt you for your API key (get one from the One dashboard) and configure the MCP server for your environment: Claude Desktop, Cursor, Claude Code, Windsurf, or any MCP-compatible agent.

Capabilities

• 1000+ platforms. Gmail, Slack, Shopify, HubSpot, Stripe, Linear, QuickBooks, and more.
• Natural language execution. "read my last gmail email", "send a message to #general on Slack"
• Code generation. "build a form to send emails using Gmail", "create a dashboard that lists my Linear projects"
• No tool bloat. Only 3 tools exposed regardless of how many platforms or actions you connect. One call finds every action a task needs, across platforms, with its documentation, so your agent's context window stays clean.
• Fine-grained access control. Restrict which actions, connections, and permission levels (read, write, admin) your agent has access to.
• Secure by default. All requests proxied through One, secrets automatically redacted, no platform API keys to manage.

Examples

Execute actions directly:

"Get my last 5 emails from Gmail"

"Send a Slack message to #general: 'Meeting in 10 minutes'"

"Get all products from my Shopify store"

Generate integration code:

"Create a React form component that sends emails using Gmail"

"Build a dashboard that displays Linear users and their assigned projects with filtering"

"Create a paginatable table that fetches and displays QuickBooks invoices with search and sort"

Tools

The server exposes three MCP tools:

Finding actions

find_one_actions takes one requests entry per operation, each a kebab-case platform and a short intent naming the operation alone ("send a message to a channel", not the message). A task describes the whole job in general terms, and an ai_model names the model asking; agents are asked to send both with every find, since the task helps tell similar actions apart. For each intent, One's decision model picks the action to use; the answer documents it, sets apart a substitute when the model was unsure, and lists a few alternatives. A platform no active platform has (clickup for click-up) is not searched: the answer names the platforms it may have meant, closest first, so the agent can ask again with the right one. Large documents come back as a digest; call find_one_actions again with load: [{ action_id, section }] for a section it left out, full: true for the whole document, or toc: true for its table of contents. ONE_ACTION_IDS, ONE_PERMISSIONS and ONE_CONNECTION_KEYS apply to every answer: an action they refuse is never offered, and when they refuse the model's pick, the next allowed candidate takes its place.

Upgrading from 1.x: search_one_platform_actions and get_one_action_knowledge are removed. find_one_actions replaces both.

Remote MCP Server

Prefer not to run anything locally? One hosts a remote MCP server at `https://mcp.withone.ai/mcp`. Point any MCP client that supports remote (HTTP) servers at that URL and authenticate with One via OAuth. There's no npm install and no ONE_SECRET to manage. You approve access in One's consent screen, where you can scope exactly which connections, actions, and permission levels the agent gets. Those choices are surfaced back to the agent through each connection's access field, so it knows what it can run without searching.

Add it to your client's MCP configuration. The endpoint speaks the Streamable HTTP transport:

{

"mcpServers": { "one": { "type": "http", "url": "https://mcp.withone.ai/mcp" } } }

Some clients omit type and take the URL alone; UI-based clients (custom connectors) just need the URL itself. On first connect, your client opens a browser to authenticate and authorize with One. After that, the same three tools are available.

One-click install:

The Windsurf link opens One's page in the editor's MCP marketplace, which is served from the official MCP registry where One is published as ai.withone/mcp. Windsurf handles the OAuth sign-in itself once you install from that page.

Manual Installation

If you prefer to configure the server manually instead of using one init, install the package directly:

npm install @withone/mcp

Then set the required environment variable:

ONE_SECRET=your-one-secret-key

Identity Scoping

Scope connections to a specific identity (e.g., a user, team, or organization):

ONE_IDENTITY=user_123

ONE_IDENTITY_TYPE=user

When set, the MCP server will only return connections associated with the specified identity, and execute_one_action will only run on those connections. This is useful for multi-tenant applications where you want to scope integrations to specific users or entities.

Access Control

Fine-tune what the MCP server can see and do:

ONE_PERMISSIONS=read

ONE_APPROVAL=write ONE_CONNECTION_KEYS=conn_key_1,conn_key_2 ONE_ACTION_IDS=action_id_1,action_id_2 ONE_KNOWLEDGE_AGENT=true

All defaults preserve current behavior. If no access control env vars are set, the server starts with full access and all tools available.

Whatever you configure here is surfaced back to the agent: list_one_integrations stamps each connection with an access field so the agent knows up front what it can run there, without spending a turn searching. It is one of:

Approval

An agent that has just read an email, a web page or another third-party document can be steered by what it read. With ONE_APPROVAL set, the person at the MCP client decides whether a change goes ahead, not the agent.

Before execute_one_action runs an action that needs approval, the server shows a prompt through MCP elicitation. It names the action, the platform, the connection and what will be sent, and the user ticks a box to run it. Declining, dismissing, leaving the box unticked or not answering within 10 minutes stops the action, and nothing is sent.

Your MCP client has to support elicitation. If it doesn't, every action that needs approval is refused, so check your client before turning this on.

Manual Configuration

If you used one init, the configuration below is already done for you. These examples are for reference or manual setups.

Standalone

npx @withone/mcp

Requires Node.js 22.23.4+, 24.22.1+ or 26.12.1+. Older Node trusts the compromised eMudhra CA certificates (October 2026), so a network attacker could intercept ONE_SECRET. On an older Node the server still starts, but every tool returns an error with the command that updates Node for how it was installed. If you can't update yet, ONE_ALLOW_UNPATCHED_NODE=1 lets it call One anyway; you remain exposed while it is set.

Claude Desktop / Cursor

Add the following to your MCP config:

• Claude Desktop: MacOS: ~/Library/Application\ Support/Claude/claude_desktop_config.json · Windows: %APPDATA%/Claude/claude_desktop_config.json
• Cursor: Open the Cursor menu and select "MCP Settings"

{

"mcpServers": {