The #1 tool for everyday use: Claude.

"Claude Code can do everything" sounds like a LinkedIn post, so here is the claim itemized instead. On my machine, this month, it has: written and published articles to this site, built the pages you are reading, organized a year of downloads into folders, pulled 3 CSV exports into one report with charts, resized and compressed a batch of images for web, drafted a week of social posts and scheduled them, checked competitor pricing pages on a schedule, and fixed its own mistakes when I pointed at them. Each of those jobs has a SaaS product charging $10 to $49 a month to do it worse.

You do not need to be a developer, and the terminal is not a skill. It is a text box. You type what you want in plain English, the agent asks before it touches anything, and you approve or decline. That is the entire loop. If the word terminal already sounds like someone else's job, keep reading. This guide assumes exactly nothing.

Same price, $20 a month each, so the honest question is not which chat is smarter this week. It is which one comes with hands. Claude Pro includes Claude Code, the version that reads and writes real files on your machine and runs real commands. ChatGPT Plus does not ship anything equivalent to a founder's laptop, and that single difference is the whole call: one of them finishes the job, the other one describes it beautifully.

Where ChatGPT still earns its keep: image generation, voice mode you can talk to while driving, and the sheer size of its plugin and GPT ecosystem. If your day is images and voice notes, keep it. If your day is documents, spreadsheets, sites, and admin that never ends, Claude is the one to pay for. I run both, but only one of them has a card on file (the free ChatGPT tier covers the image jobs Canva does not).

The rule I actually use: pay for the assistant that can act, use the free tier of the one that only talks. That is the AI Chat row of the Stack, settled.

Open Terminal on a Mac (Cmd+Space, type "terminal") or PowerShell on Windows. Paste the line for your machine and hit enter:

# Mac and Linux
curl -fsSL https://claude.ai/install.sh | bash

# Windows (PowerShell)
irm https://claude.ai/install.ps1 | iex

That is the official installer. It keeps itself updated and needs nothing installed first. If you already live in Homebrew or npm, brew install --cask claude-code and npm install -g @anthropic-ai/claude-code both work too. When it finishes, type claude in any folder and you are in. (Verified July 2026.)

First run of claude opens a browser and logs you in with the Claude account you already have. Then the plan question, answered plainly:

Pro, $20 a month ($17 on annual) includes Claude Code and covers a normal solo-founder day. Start here, always. Max, $100 or $200 a month, is 5x or 20x the usage: buy it the second week you hit Pro's limit twice, not before. API billing (pay per token, no subscription) is for automations that run while you sleep, and the section on tokens below covers exactly when that flips. Prices verified July 2026.

Here is the part the pricing page will not do for you. Write down what the agent takes over in month 1, then cancel those:

• The automation middleman, $20 to $30 a month. One agent plus the connections in the MCP section below does the same shuttling.
• The AI writing seat, $29 to $49 a month. Almost always a prompt with a pricing page attached.
• The single-job utilities, $8 to $15 each. File cleanup, bulk image resizing, PDF wrangling, CSV merging. I had 4 of these.
• The reporting add-on your other tools charge for, $15 and up.

That is $70 to $120 a month against one $20 plan, which is the only reason I can run this business on 22 tools instead of 40. Do not cancel on faith, though. Cancel the week after the agent has actually done that job twice.

Claude Code starts cautious. It asks before editing a file or running a command, and you should leave it that way for week 1, because approving each step is how you learn what it actually does. Shift+Tab cycles the permission modes: auto-accept edits, or plan mode, where it proposes an approach and touches nothing until you agree. Plan mode is the underrated one. Big jobs go wrong in the plan, not the typing. Auto-accept is the one to earn: leave it off until you have a defined process you have watched work, because a faster agent on an undefined process is just a faster way to be wrong.

Then give it a memory. Run /init inside your project folder and it writes CLAUDE.md, a plain text file the agent reads at the start of every session. Edit it like you are onboarding a hire: what the business is, how you like things written, what it should never touch. Keep it under a couple hundred lines. This file is the difference between a tool and a colleague.

Do not build a demo app. Give it a job you were about to pay for:

The file mess. "Look at my Downloads folder and organize everything into folders by type and year." There goes the $10 cleanup utility.

The report. "Read these 3 CSV exports and build one spreadsheet with monthly totals and a chart." There goes the reporting add-on.

The images. "Resize every image in this folder to 1080px wide and compress them for web." There goes the bulk-image SaaS.

Each one finishes inside 10 minutes, and each one is a tool that never gets to bill you. When you catch yourself googling "best app for X", stop and ask Claude Code first. That reflex is the product. The tools that survive the reflex are on the Stack.

The people who get the most out of Claude Code stop treating it like autocomplete and start treating it like a hire. 3 habits carry most of the value:

Brief the job, not the keystrokes. "Here is the goal, here are the files, here is what done looks like" beats 20 micro-instructions. One full brief costs less and lands better than 5 corrections, because every correction re-reads everything before it.

One session per job. Finish the newsletter, type /clear, start the site fix fresh. Old context is pure tax on the new job, and mixed jobs produce mixed work.

Make it verify. End real jobs with "now check it": open the page, run the script, read the output back. An agent that checks its own work is the difference between delegating and hoping.

And when a job repeats weekly, tell it so. Claude Code can write itself a checklist file for the routine, and next week the whole job is one sentence. That is the moment the $29-a-month automation aisle quietly closes.

Every integration platform runs the same pitch: pay monthly to shuttle data between tools you already pay for. Zapier from $30, Make, the "connector" add-on on every SaaS pricing page. That pitch had a good decade. It is over, and the reason is 2 words most people use interchangeably when they are not the same thing at all.

Connectors and MCP are not the same thing

Connectors are the one-click list inside Claude. Open your settings, find Connectors, browse, click, log in once, done. No terminal. It is a good list and it is a short list.

MCP (Model Context Protocol) is the open standard underneath the whole thing. It is how any tool makes itself available to Claude. The connectors list is just the pre-built shelf, and most of what you actually run every day is not on that shelf. Those you add yourself as a custom connector by pasting in the tool's MCP address.

That is the entire difference: the shelf, and everything else. Almost every guide stops at the shelf, which is why people conclude this stuff is not ready. The everything else is where the hours are.

The 5 connectors worth switching on today

Google Drive. It reads and writes docs and sheets directly, so a plan lands in the doc instead of dying in a chat window. If you connect one thing this week, connect this one. It takes 30 seconds.

Canva. Hand it the content, it pushes it into a template. This is the whole reason a carousel a day is possible without a designer.

Gmail. "What did that supplier quote me in June" without opening a single thread.

Google Calendar. The boring one that saves the most, because an agent that can see your week stops proposing things you have no room for.

Stripe. Revenue questions in plain English. "What actually sold last month" without building a report.

The 3 MCPs that are not on the shelf

These are the ones that changed the week, and not one of them is in the directory. You add them yourself.

Metricool. It schedules the posts to every platform. The gap between "the content is written" and "the content is posted" used to eat an afternoon a week. That gap is gone. See the scheduler review for why it is the pick.

Flodesk. It wires the email subscribe form natively into a new site, so the site collects addresses on day 1 instead of week 3, and it reads the campaign numbers back to you.

Higgsfield. It makes the visuals from a written brief, which is the half an agent genuinely cannot do itself.

How to check what you have wired

1. Open Claude, go to Settings, find Connectors. Everything already connected is listed there.

2. Browse the directory in the same place and switch on the ones you already pay for.

3. For anything not in that list, use Add custom connector and paste the tool's MCP address. Searching "[your tool] MCP" usually lands on it in their own docs.

4. In Claude Code, type /mcp inside a session to see what is wired and log into anything that needs it. To add one from the terminal instead:

claude mcp add --transport http <name> <server url>

claude mcp list        # see everything connected

Servers are added per-project by default; --scope user makes one follow you everywhere. (Verified August 2026.)

3 is the right number for month 1. If a connection is not saving you a login or a subscription within 2 weeks, remove it. And one honesty note before you plug anything in: whatever you connect, the agent can act inside. Give it the trust you would give someone in their first week, not their fifth year. Prefer read-only where it is offered, and skip servers from vendors you have never heard of. The point of this page is fewer tools you trust more, not a new collection.

Every AI tool bills in tokens now, and tokens are where the old subscription creep went to hide. On a Claude plan you feel it as hitting your usage limit at 3pm. On the API you feel it as a bill. The cure is the same either way: context discipline. Claude carries the whole conversation with it on every turn, so the longer and messier a session gets, the more every single message costs. Keep the context small and the spend follows.

3 ways to use it

1. One chat per workflow. Keep them separate. A thread that has already done 3 unrelated jobs makes the fourth one expensive, because all of it comes along for the ride. Finished the newsletter and moving to the site? /clear, or just start fresh.

2. Skills and CLAUDE.md. Write your documentation down as files the agent reads, so you are not re-teaching it who you are and how you work at the start of every session. A skill is a folder with instructions it loads only when the task calls for it, which is the cheap way to hold a lot of context. Keep CLAUDE.md itself lean, though: it loads every single session, so 200 lines is memory and 800 is a standing charge.

3. Point at files, do not paste them. Copy the filepath and hand it that. Pasting 400 lines means you pay for those 400 lines again on every follow-up question, forever. And /context shows exactly what is filling the window when you want to know where it went.

3 ways to cut what you burn

1. Use the right model. Simple prompt, a bit of text, a rewrite? Sonnet. Problem solving and anything genuinely complex? Opus 5. Agents running something hard is where Fable earns it. Running the top model on a job that did not need it is the easiest way to burn a day's usage before lunch. Switch any time with /model.

2. Do not compact the chat. This is the one that goes against the grain. Compacting summarizes and carries on, which keeps you inside a thread that is already heavy and quietly drops detail you will want later. The better move: have it update your documentation, then ask it to write you a continuation prompt for the next task, and open a fresh chat with that. A handoff note plus a clean session costs a fraction and you lose nothing.

3. Make it confirm the plan before it executes. The big one. Left alone it assumes what you meant and goes and does it, cheerfully building you a PDF nobody asked for, and that is real usage spent on something that was never the goal. Tell it to check the plan with you first. Shift+Tab into plan mode does the same thing structurally. You feel the difference within a day.

Then, and only then: once you have a defined system you know works, turn auto on. Not before. Auto-accept on an undefined process is just a faster way to be wrong.

When you want receipts: /usage shows where your plan limits stand, /cost shows API spend. (Verified August 2026.)

Which model, when

*Sonnet 5 runs $2 in and $10 out on introductory pricing until September 2026.

On a subscription you never see these dollar figures, but they still run your day: pricier models burn your usage limit faster. The rule that keeps the bill boring is to run the cheap model by default and escalate when the job resists. Drafting, cleanups, and summaries do not need the top tier. Debugging a gnarly automation at midnight does. Switch any time with /model.

Spend-shaping moves

Batch the ask. One message with the full spec beats 5 corrections, because every correction re-reads everything before it. The 3-step sequences in Prompts are written this way on purpose.

Send the heavy reading to a subagent. For big research or codebase questions, ask Claude Code to use a subagent: the digging happens in a side context and only the conclusion comes back to yours.

One-shot the small stuff. claude -p "your question" answers and exits. No session, no accumulating context, no tax.

A page this long owes you the other side. 4 jobs I do not hand it:

Images and video. It writes the brief, it does not make the asset. Design still goes to Canva and generated video still goes to Higgsfield. Anyone selling you "one AI for everything visual" is selling a wrapper.

Publishing to social. It can draft and schedule through a tool, but the queue, the previews, and the analytics belong to a real scheduler. I tried the agent-only version for a month and lost posts.

Anything unattended without a budget. Long jobs that run while you sleep want API billing and a spend cap. On a plan you will simply wake up rate-limited.

Confident wrong answers. It will state a number with total composure. Anything that touches money, tax, or a public claim gets checked by a human, which is me, which is why every review here says "verified" with a month attached.

None of that changes the call. It changes the shape of the stack around it: one agent doing the work, and the small number of tools it genuinely cannot replace.