How to Connect Your Tools to an AI Agent: OAuth, API Keys and MCP

Most AI agents connect to tools in one of three ways. OAuth lets you approve access on the tool's own sign-in screen. An API key is a secret you paste in. MCP is an open standard that lets AI assistants reach other apps. Prefer OAuth where it's offered, since you share no secret and can revoke access.

An AI agent is only as useful as what it can reach. Connections also carry the most risk, because each one is access to your business. This guide explains the three methods in plain words, then shows how to connect tools without giving away more than the job needs.

These aren't strict alternatives. MCP describes how an assistant talks to a tool, and it often uses OAuth for the sign-in. That's how it works with Chargezen.

Use it whenever it's offered. You sign in on the tool's own page, see what's being requested and approve it. No password or long-lived secret changes hands in your chat or your files. In Concierge, services that use a sign-in show an Authorize button on the Integrations page, and Reauthorize if the access expires.

Read the permission screen. If a reporting agent asks to send and delete mail, ask why.

Some services only offer keys, and they're common for AI model providers. If you use a key, follow these rules:

• Create a separate key for each purpose, so you can delete one without breaking others.
• Limit the key's permissions if the service allows it.
• Never paste it into chat, and keep it out of documents and code.
• Enter it only in a secret field in the product's settings.
• Replace it when someone with access leaves.
• Set spending limits with the provider where you can.

In Concierge, you enter keys in secret fields on the Integrations page and can disconnect them at any time. One common use is to run agents on your own model account. Calls on your key are billed by your provider rather than from your credits. See bring your own key.

Model Context Protocol (MCP) is an open standard that lets AI assistants reach other apps through one common method, instead of a custom connection for each pair. It matters most when your team already works in an assistant like ChatGPT or Claude and wants to use business tools from there.

Here's how it works with Chargezen. You add the Chargezen MCP server to your assistant, sign in, pick a workspace and review what the assistant may do. There's no API key to copy. The assistant then acts as you, with your permissions, and it can look up orders, pull reports or hand a job to Concierge. Chargezen checks your access on every request, so it can never do more than you can.

Changes that delete, replace or publish something ask first. Clicking Disconnect ends the assistant's access immediately, and removing a teammate from the workspace ends their assistant's access too. Treat any MCP server like any other app you install: check who runs it and what it can do.

• 1. Start with the job: Write down what the agent must read and what it must change. Connect nothing else.
• 2. Choose the method: Prefer OAuth, and use a limited API key if that's the only option.
• 3. Read what it requests: Check the permissions before you approve.
• 4. Give it to one agent: In Concierge, connecting a service doesn't hand it to every agent. You tick the capabilities on each agent's Configure Agent page.
• 5. Test with something harmless: Run a read-only request and check the result.
• 6. Add write access later: Add one change at a time, once reads are reliable.
• 7. Note who connected it: Know who owns each connection, and review them on a regular schedule.

Many supplier portals have no API. A Concierge agent can use its own browser to read those sites, limited to the sites allowed for that job. When a page needs a login, it hands you the screen and you sign in yourself. Credential handover keeps your password out of the chat and the agent's memory, and the sign-in lasts only for that job. See the computer and browser docs for what it can and can't do there.

Connecting your own Google or Microsoft account differs from connecting a shared company tool. The agent is reaching your personal mail, calendar or files, so choose narrowly. In Concierge, files, mail and calendar are separate permissions you turn on one at a time, and removing a teammate from the workspace revokes their personal grants.

• Which tools are connected, and does each one still have a job?
• Which agents can use each tool?
• Who connected each one, and do they still work for you?
• Which assistants are connected? In Concierge, see Deploy, then Channels, then External MCP servers.
• Have any keys been shared or left in old documents?

Concierge integrations are grouped on one page, with access set for each agent. ChatGPT and Claude can reach your workspace through the Chargezen MCP server, and the setup guide covers each step. All Chargezen integrations are on the integrations page.

Keep reading: an AI agent security checklist, where your data goes and AI agents and subscriptions over MCP. Glossary: OAuth, Model Context Protocol and least privilege access.

An API is the way one piece of software talks to a particular service, and each is built differently. MCP is a common standard that lets AI assistants find and use tools the same way across many apps. An MCP server often calls APIs behind the scenes.

Usually, for connecting an agent. With OAuth, you approve specific access on the tool's own screen and can revoke it, and no secret passes through chat. A key can work well too if it has limited permissions and is stored in a secret field.

No. Use OAuth, a limited API key or a handover where you sign in yourself. With Concierge, a site login happens on a separate page, and your password never goes into the conversation.

Disconnect it in the agent product and, for OAuth, remove the app in the tool's own settings too. Delete any API key it used. In Concierge, use Disconnect on the Integrations page, or on External MCP servers for an assistant.