Guide
MCP, the Model Context Protocol, is an open standard that lets an AI assistant connect to outside data and tools through small programs called servers, so you stop pasting the same context into every chat. It is worth setting up when you keep re-explaining the same files or records. Dayflow fits only if the context you want is your own work history on a Mac.
You open a chat with an AI assistant and spend the first five minutes setting the scene. You paste in the project brief. You paste in the notes from yesterday. You explain, again, who the client is and what “the March version” means. Then you ask your actual question, and halfway through the answer you realize it needs a file you forgot to include.
Meanwhile the letters MCP keep showing up in release notes and settings screens. “Now supports MCP.” “Add an MCP server.” Everyone seems to assume you know what it means, and the explanations you find are written for people who build software.
This guide explains the Model Context Protocol in plain words: what it is, why it exists, what can go wrong, and a handful of things to do before and after you connect anything. Dayflow, the Mac app that publishes this blog, comes up once near the end. Dayflow details here come from Dayflow’s own documentation for version 2.2.0, as reviewed on 2026-09-06.
An AI model on its own only knows what it learned in training and what you type into the conversation. It does not know about your files, your calendar, your team’s tracker or the notes you wrote last Tuesday. For a long time the fix was manual: copy, paste, explain.
The next step was custom connections. Each AI app built its own way to read a calendar or a folder, and each data source had to be wired up separately for each app. That works, but it multiplies quickly. Five apps and ten sources means a lot of one-off plumbing.
MCP is an attempt to replace that with one shared plug. The usual comparison is a standard charging port: if every device uses the same shape, any cable works with any device. In MCP terms there are two sides:
A server can offer three kinds of things. Resources are data the assistant can read, like a document or a list of records. Tools are actions it can take, like searching, creating a file or sending a message. Prompts are ready-made instructions the server suggests for common tasks.
Servers can run on your own computer or on someone else’s machine over the internet. That detail matters more than people expect, and the steps below come back to it.
Why it feels confusing: the word “protocol” sounds like something you install, but it is closer to an agreement about format. You never use MCP directly. You use an app that speaks it, and you add servers that speak it too.
Why it can feel risky: an assistant connected to real sources can read real data and, with tools, take real actions. Text it fetches can also contain instructions that were never meant for it, such as a line hidden in a web page or an email that tries to steer the assistant. Many people find the fix is less about the technology and more about being deliberate with what they connect.
Most of these need nothing more than paper and a week of paying attention.
For one week, keep a sheet of paper next to your keyboard. Every time you paste context into an AI chat, write one line: what you pasted and where it came from. Do not try to fix anything yet. Just count.
Illustration: by Friday the sheet might read “project brief (shared folder) x6, meeting notes (notes app) x4, yesterday’s work (memory, typed out) x5, error log x2.” The top line is your first candidate for a connection. The “yesterday’s work, from memory” line is a different problem, and step 6 and the Dayflow section below deal with it.
After a week: if one source shows up far more than the others, a single server for that source is probably worth setting up. If the tally is scattered and short, you may not need MCP at all yet, and that is a perfectly good answer.
Fold a page into three columns: Never, Ask me first, and Fine. Sort your likely sources into them. Client contracts, health information, passwords and anything covered by an agreement with your employer usually land in Never. Draft documents might be Ask me first. A public reading list might be Fine.
Then answer one more question in writing: where does your AI app process the conversation? If it is a cloud model, anything a server returns travels into that conversation. A server running on your Mac does not change that.
Illustration: someone might put “shared project folder” in Fine, “email” in Ask me first, and “client invoices” in Never, then realize the shared folder contains a subfolder of invoices and decide to point the server at a narrower path.
After a week: the page is working if you have looked at it at least once before approving something. If nothing you have connected falls outside the Fine column, the line is doing its job.
For every server you plan to add, write one sentence: “This lets the assistant ____ so that I don’t have to ____.” If you cannot finish the sentence, do not add the server.
Illustration: “This lets the assistant read the brief folder so that I don’t have to paste the brief into every chat.” Clear, narrow, easy to check. Compare that with “This connects my whole drive in case it’s useful,” which is hard to evaluate and easy to regret.
After a week: reread the sentence. Did the connection do that job? Did it do things you did not intend? If the honest answer is “I forgot it was there,” remove it.
Pick the top line from your tally and add a single server for it. Where the setup offers a choice, choose read access only. Actions that change things, such as writing files or sending messages, can wait until you trust the read side.
Illustration: a writer connects only a folder of research notes, read-only. For the first few days they ask questions like “what did I already note about this topic?” and nothing else. Once that feels routine, they consider whether anything else is worth adding.
After a week: count how many times you would have pasted from that source and did not have to. If that number is meaningful to you, the connection is earning its place. If you kept pasting anyway out of habit, the problem may be the habit, or the connection may not reach what you actually need.
When you add a server, most AI apps show what it offers: a list of resources and tools with short descriptions. Read it the way you would read app permissions on a phone. Look for anything that writes, deletes, sends or spends.
Many clients also ask for approval before a tool runs. Keep that turned on, at least for actions that change things.
Illustration: a server described as “read your calendar” turns out to include a tool called “create event” and another called “invite attendees.” The person keeps the server but declines those two tools when the assistant asks, and makes a note on their page from step 2.
After a week: the habit is working if you can name, from memory, what each connected server is able to do. If you cannot, the list is too long or you approved too quickly.
At the end of each day, write three lines: what you asked the assistant, what it fetched, and whether the answer was right. A plain notebook works. So does the daily work log template if you already keep one.
Illustration: “Asked for a summary of open questions on the redesign. It read the brief folder and the meeting notes. Missed the decision from Wednesday’s call because nobody wrote it down.” The last part tells you the gap is in your notes, not in the connection.
After a week: look for patterns. If the assistant keeps fetching the wrong thing, the server is pointed too broadly. If answers keep missing the same kind of information, that information probably is not written down anywhere, and no protocol can retrieve what was never recorded.
One gap shows up in many tallies: “what I worked on yesterday, typed out from memory.” That information usually lives nowhere. You reconstruct it, and reconstructions are fuzzy, which is the whole subject of our page on where your day went.
Dayflow is a free, MIT-licensed Mac app (macOS 14 or later) that records the screen at one frame every 10 seconds and has AI write the day into a readable timeline. Full recordings and the timeline are stored on your Mac, so there is no timer to remember to start.
You choose how screenshots are analyzed. Local AI runs a supported model on your Mac through Ollama or LM Studio, and screenshot analysis stays on-device. You can instead send selected screenshots to a cloud provider you configure, under that provider’s terms, or use Dayflow Pro at $20 a month, which manages cloud analysis. The local and cloud processing page covers each boundary.
On assistants: an assistant can receive the Dayflow data it retrieves in your conversation, and local storage does not make a cloud assistant conversation local. That is the same rule from step 2. Supported setups are on the integrations page. If you would rather hand over a file, you can export a date range as Markdown from Settings → Export, as described in the export workflow.
AI summaries can miss or misinterpret work, so check them before sharing. Dayflow publishes this article.
Be clear about what it is and is not.
MCP is a shared plug between AI apps and the places your information lives. It does not make an assistant smarter. It makes the assistant better informed, and only about the sources you choose to connect.
The useful work happens before you install anything. Count what you keep pasting. Decide what should never leave your hands. Connect one source, read-only, with a one-sentence job. Read what each server can do, and keep a short log for a week. If the log shows you saved real effort, add the next connection. If it shows the missing piece was never written down, the fix is a better record of your work, and that is a smaller, quieter problem than it sounds.
MCP stands for Model Context Protocol. It is an open standard that describes how an AI app can connect to outside sources of information and actions. The source runs as a small program called a server, and the AI app acts as the client. Because the format is shared, one server can work with any client that supports the protocol, instead of each app needing its own custom connection.
MCP is a format for connecting things, so safety depends on which servers you add and what they can reach. A server that reads one folder is a small risk. A server that can send messages or delete files is a bigger one. Add servers from sources you trust, read the list of actions each one offers, start read-only, and approve actions one at a time where your app allows it.
Not always. Some AI apps let you add a connection through a settings screen, with no code. Others still ask you to edit a small configuration file and paste in a command. You do not need to write a server yourself to use one. If a setup guide asks you to run commands you do not understand, it is reasonable to stop and ask someone who does.
The client is the AI app you talk to. The server is a separate program that sits next to a source, such as a folder, a database or a calendar, and offers that source to the client in a standard way. The client asks, the server answers. One client can connect to several servers, and one server can be used by different clients that support MCP.
MCP itself does not decide that. A server can run entirely on your computer, but whatever it returns goes into your AI conversation. If that conversation is handled by a cloud model, the returned data leaves your machine as part of the chat. If the model runs locally, it may not. Check where your AI app processes conversations before connecting anything sensitive.
Not quite. An API is the way one specific service lets programs talk to it, and every API looks different. MCP is a shared layer on top: a server often wraps an existing API and presents it in one standard shape that AI apps already understand. The benefit is that an AI app only has to learn MCP once instead of learning every service's API separately.
Dayflow's published facts say an assistant can receive the Dayflow data it retrieves in your conversation, and that local storage does not make a cloud assistant conversation local. Which assistants are supported and how to set them up is listed on Dayflow's integrations page. You can also export a date range as Markdown from Settings, then Export, and share only that file.
No. Dayflow records what was on your Mac's screen and turns it into a timeline of your day. If you want an assistant to read a project folder, a set of documents or a codebase, a server built for that source is the better pick, because it hands over the actual files rather than a summary of what you looked at. Use Dayflow when the question is about your time.
No. Dayflow does not provide website blocking. It records your screen and writes a timeline after the fact, so it can show you where an afternoon went but it will not stop you from opening a site. If blocking is what you need, a dedicated website blocker is the right tool, and you can use it alongside a record like Dayflow if you want both.
Dayflow is $0 and open source. Full recordings stay on your Mac, with local AI available for on-device analysis.
Download for MacOS macOS 14+ · MIT licensed · no account required