How to Connect Any API to Claude, GPT, or Gemini via MCP

“Can my AI agent just check our CRM?” is one of the most common questions teams ask once they start using Claude or another AI assistant for real work. The honest answer used to be: only if someone writes custom integration code for that exact use case. MCP changes that answer to: yes, in minutes, and it stays working for every future use case too.

Here is what actually connecting an API to an AI agent looks like in practice — whether it's a well-known SaaS tool or something you built yourself.

Step 1 — decide what your agent needs to touch

Start narrow. Pick one real, recurring question your team asks that requires checking an external system — “what's the status of this deal in Salesforce,” “are there open issues on this repo,” “has this invoice been paid.” That single use case is enough to prove the pattern before you expand it.

Step 2 — check the catalogue before you build anything

Common tools already have pre-built MCP connectors: SAP, Salesforce, HubSpot, GitHub, Jira, Slack, PostgreSQL, Stripe, and more. If the API you need is already in a catalogue, this step is just entering credentials — no code to write.

Step 3 — for anything else, bring your own API

Internal tools and niche APIs won't be in any catalogue. For those, there are two paths:

• →Upload an OpenAPI spec — the platform reads it and generates MCP tool definitions automatically
• →Define tools manually — name, HTTP method, endpoint path, parameters, and which fields to pull out of the response

Either way, the result is the same: a tool definition your AI agent can discover and call, with a clear description of what it does and what inputs it expects.

Step 4 — paste one endpoint into your agent

This is the part that surprises people: you don't configure each connector separately inside your AI agent. You configure one MCP endpoint, once, and every connector you've activated becomes available through it.

{ "mcpServers": { "yourplatform": { "url": "https://mcp.yourplatform.com/you/sse", "headers": { "Authorization": "Bearer your-key" } } } }

Add a fifth connector next month, and it shows up on the same endpoint automatically — no change needed in Claude Desktop, Cursor, or wherever your agent lives.

Step 5 — ask the question

From here, it's just conversation. The agent reads the available tools, decides which one answers the question, calls it, and replies in plain language:

• →sap_get_purchase_order({ po: "4500012345" })
• →github_list_issues({ repo: "acme/backend" })
• →salesforce_get_deal({ id: "006Hs00001xyz" })
• →stripe_get_invoice({ id: "in_1Nxxxxxx" })

What can go wrong (and how to avoid it)

The two failure modes worth planning for from day one: credentials leaking, and not knowing what your agent actually did. Store every credential in an encrypted, per-account vault rather than plain environment variables shared across connectors, and keep a log of every tool call — input, output, duration, and status — so any unexpected agent behaviour is traceable after the fact rather than a mystery.