16. MCPs & APIs
Note: The video covers material not in the guide below — please watch in full.
Action Step
Complete this before moving on.
Follow along with the video and connect Claude Code to both Fireflies and Teamwork. For each one, go to the platform's settings, copy your API key, and tell Claude to connect to the API. Once connected, have Claude create an API skill for each platform so future sessions can reconnect without redoing the setup. Then try pulling real data — your recent transcripts from Fireflies or your tasks from Teamwork.
Training Guide
So far, everything your AI has done has been local — reading files on your computer, writing documents, searching through folders. Powerful, but limited to what's already on your machine.
What if you could ask your agent: "What tasks are due this week in Teamwork?" and it just... pulls the answer? No logging in, no navigating menus, no copy-pasting. Just ask and get.
That's what this training unlocks.
(But first, let's understand the concept behind it)
What's an API?
You've probably heard this word before. API — Application Programming Interface. Sounds technical. Here's what it actually is:
An API is how software talks to other software.
Every platform you use — Teamwork, Slack, Google Calendar, your bank — has an API. It's a door that other programs can knock on to ask for data or perform actions. Your budgeting app doesn't log into your bank's website. It knocks on the API door, asks "what's the balance?" and gets a clean answer back.
The catch? Traditionally, you need a developer to use an API. Someone writes code, handles authentication, parses responses, deals with errors. So you just use the platforms directly — open the browser, log in, click around.
Then AI came out and everything changed.
(First came MCPs)
MCPs — Pre-Built Connections
In November 2024, Anthropic launched the Model Context Protocol — MCP for short. Instead of needing a developer to connect to an API, you give the AI a pre-built connection. A config file that says "here's how to reach Teamwork, here's the authentication, here's what you can do." The AI reads the config and the connection just works.
MCPs took off fast. By March 2025, OpenAI adopted the same standard — choosing MCP over their own plugin system. Google and Microsoft followed. Just like skills, MCPs became an industry-wide standard. Anthropic donated it to the Linux Foundation, and today there are thousands of pre-built connections available.
Why did they explode? Timing. When MCPs launched, the models weren't reliable enough to write raw API code on their own. They'd hallucinate endpoints, botch authentication, misformat requests. MCPs solved that — pre-built, tested, reliable. Right solution, right time.
(Then something bigger happened)
The Agents Got Smart Enough
The models kept improving. And at some point, they got good enough to write API code themselves, in real time.
Claude Code is the developer now. You say "pull my tasks from Teamwork" and it writes the API call on the spot — handles authentication, formats the request, parses the response — and hands you the data. You don't even see the code. It happens behind the scenes.
No developer. No pre-built connection needed. Just the agent talking to another system because you asked it to.
This means you're not limited to platforms that have an MCP. Any system with an API — and almost every modern platform has one — is now accessible through your agent.
(So which one do you actually use?)
MCPs vs Direct API Access — Where Things Stand
Both work. Here's the practical difference:
MCPs save you tokens. Writing API code from scratch eats context window. The agent has to figure out authentication, endpoints, data formats — all of that costs tokens. An MCP skips that. Pre-built connection, straight to the data.
Direct API access gives you reach. There are thousands of MCPs, but there are APIs for everything. No MCP for the tool you need? Claude connects directly. You're not limited to what someone has pre-built.
The practical rule: MCP exists? Use it — faster, cheaper on tokens. No MCP? No problem — Claude hits the API directly.
The line between the two keeps blurring. As models get smarter and context windows grow, we may not need pre-built connections at all. But right now, MCPs are a useful optimization for tools you use every day.
(Let's see this in action)
Guided Demo: Connecting to Fireflies
Fireflies is the tool that records and transcribes your meetings. Let's connect Claude to it — directly through the API.
Step 1: Get your API key
- Go to Fireflies → Settings → Developer Settings
- Find your API key and copy it
Step 2: Ask Claude to connect
"I want to pull my recent meeting transcripts from Fireflies. Here's my API key. Can you connect to their API and get my last 5 meetings?"
Claude looks up the Fireflies API, writes the code to authenticate, makes the request, and pulls your transcripts. You just see the results.
Step 3: Try something useful
"Find the transcript from my meeting with [Client Name] last week and summarize the key action items."
"Pull all my meetings from the last 30 days and list who attended each one."
No developer. No pre-built connection. Claude talked directly to the API.
(Same thing works for any platform — let's do one more)
Guided Demo: Connecting to Teamwork
Teamwork is the project management tool the team uses daily. Same process.
Step 1: Get your API key
- Log into Teamwork
- Go to your profile settings → API section
- Generate or copy your API key
Step 2: Ask Claude to connect
"I want to pull my project data from Teamwork. Here's my API key. Can you connect to their API and list my projects?"
If you see your projects — you're live. Now try something more useful:
"What tasks are due this week across all my Teamwork projects?"
"Give me a status summary of the [Project Name] project."
"Who has the most tasks assigned in the Marketing project?"
Pulled directly from Teamwork. No browser. No clicking. No spreadsheet. Same pattern as Fireflies — same pattern as anything.
(This changes how you think about platforms entirely)
"I Don't Like to Log Into Platforms Anymore"
Yasin's Joke: I don't want to use Chrome anymore and I don't want to click around in an app anymore. ANY website (notion, Calendar, TeamWork - just have my agent connect to it and it'll do all the work!")
At work:
- Project status? "What tasks are due this week in Teamwork?" — no logging in, no clicking through projects
- Meeting prep? "Pull the transcript from my last call with NovaPay and summarize the action items from Fireflies"
- Slack catch-up? "What did the marketing team discuss in Slack while I was out yesterday?"
- CRM update? "Log a note on the NovaPay deal in HubSpot — met with their VP, they want a proposal by Friday"
- Email draft? "Check my last 3 emails from Sarah in Gmail and draft a reply to her latest question"
- Reporting? "Pull this month's pipeline numbers from Salesforce and format them for the team update"
Personal:
- Finances? "What did I spend on restaurants last month?" — pulled straight from your bank's API
- Calendar? "What's my schedule look like tomorrow? Move my 2pm to 3pm"
- Travel? "Find me flights to Miami next weekend under $300"
- Smart home? "Turn off the living room lights and set the thermostat to 68"
- Music? "Play my Discover Weekly playlist on Spotify"
The pattern is always the same: instead of opening an app, navigating menus, and clicking buttons — you just ask. The platforms are still there. You're just accessing them through a faster door.
(One more thing — finding new connections)
How to Connect to Anything
You don't need to memorize anything. Two paths, same starting point:
Ask Claude.
If you want a pre-built connection (MCP):
"Is there an MCP or API for [platform name]? Help me set it up."
Claude finds it, walks you through setup, tests the connection.
If you want to connect directly to an API:
"I need to pull data from [platform]. Can you connect to their API?"
Claude researches the API, writes the connection code, and gets you access.
Either way — you're in. First connection takes a few minutes. After that, it's routine.
(This was the last of the eight abilities we previewed at the start — permission modes, queuing, file system access, tokens, sub-agents, skills, and now MCPs and APIs. You've experienced all of them. Next up: putting it all together)
Comment in Slack
Post your answer in your onboarding channel.
What was your biggest takeaway(s) from this training?