This was originally published on All Agents Considered.
Tomorrow, Google will switch off a free tool a lot of people built their daily work around. Most readers here have never touched it, so the shutdown itself won't ruin your week. What sits underneath the shutdown is the part worth your attention. It keeps happening, and one day it might land on a tool you do depend on.
Gemini CLI is a small terminal tool Google released to run AI agents straight from your computer. Think of it as Google's answer to Claude Code or Codex, a way to point an agent at your files and let it work.
What made it interesting was that it shipped open source, which means anyone was free to read the code, copy it, fix it, or build something better on top of it. Open source is the closest thing software has to a promise that a tool stays yours even if the company behind it loses interest.
Google made that promise, and now Google is walking away from it.
Last month, Google announced that on June 18 2026, the tool will stop working with Google's AI subscription plans. In its place comes something called Antigravity CLI, which is closed-source.
Six days earlier, the US government forced Anthropic to shut down Fable 5, a model people had barely started using before it disappeared.
Two different shutdowns, same lesson. The tool you built your work around can change or disappear, and you won't get a vote.
That's why I built my AI stack so I can always swap models without rebuilding anything. When one company shuts the door, another one is already wired up and ready to go. In today's edition, I am going to show you exactly how to break free of vendor lock-in.
In this piece:
- Why three different AI vendors made three different moves in one month, and how every move pointed at the same trap
- How to spot the parts of your own AI setup that secretly hold your context, your memory, and your routines hostage
- The three decisions I built mine around so any single vendor change turns into a shrug instead of a crisis
- One move you finish in twenty minutes today that proves the whole principle to yourself
A Pattern Everyone Recognizes and Nobody Names
Here's how it goes, almost every time. A company hands out something free, often open source, and people build real work around it because free and open feels safe. Word spreads. More people lean on it.
Then a quiet thing happens inside the company. Someone notices the free tool now does roughly what the paid product does, and the two start eating each other. Around that point the terms change, or the license shifts, or the whole tool gets retired and replaced by a sealed version.
Goodwill came from the open thing. Money comes from the closed thing. When those two collide, money wins every single time.
Gemini CLI walked this exact road. It carried an Apache 2.0 license, which in plain terms meant stays open forever, free for anyone to keep using even if Google lost interest.
Last month's announcement erased that. A permanently open tool moved behind a paywall and into closed source. A community of people who trusted the open license now have less than 48 hours to migrate or go dark.
Anthropic ran a version of the same play a week before Google's announcement. They quietly walked back hidden safeguards inside Claude after a public backlash.
People leaning on Claude for serious research found out their tool had been working against them in the background. It shifted its behavior while they changed nothing about how they used it.
Then the US government forced Anthropic to shut down Fable 5 entirely. People who had just started building on it lost access overnight, and the company had no choice but to comply.
Two different moves from the same company in the same week. One changed the tool's behavior in secret. The other made it disappear completely.
Same week, OpenAI floated price cuts aimed squarely at Anthropic. Your monthly AI bill now rides on a boardroom argument you'll never sit in and never hear about.
Three companies, three different moves, one shared lesson. Lean your whole setup on a single vendor's tool or model, and you take on their shutdown dates, their hidden behavior changes, their price wars, and every private decision about which features live and which ones disappear overnight.
What Dependency Looks Like
Most people never feel dependent until the morning something breaks. You sit down with coffee, open Claude, and a prompt it handled fine yesterday gets refused today for no reason you understand.
In that moment you learn your entire research routine was balancing on one model staying agreeable. Or you open ChatGPT to pick up a three-month project and your conversation history is gone, wiped during a policy change you never read.
A tool you built your whole day around changed its mind. Your only role in the decision was finding out afterward.
Dependency runs deeper than the model itself. It reaches into everything the model has been touching on your behalf.
Your project context lives inside a chat window that vanishes the second you close the tab. Your task history sits trapped behind an interface with no real export. Your routines and saved instructions live in one company's private format.
Moving to a competitor later means rebuilding the whole thing by hand from memory. Nobody handed you a contract to sign. You used a tool that felt good in the moment.
The context you kept adding to it became the most valuable thing in the room, right up until the day it walked out the door wearing the tool's logo.
This was my normal life before I built around it. I would spend an hour walking Claude through a project, finally get answers worth keeping, close the laptop, come back the next morning, and start from zero.
Every time I switched providers chasing a better price or a smarter model, every routine reset to nothing. None of it hurt enough to fix on any single day.
It only became unbearable after it happened enough times that building a real fix took less energy than complaining about it one more morning.
Three Decisions That Make Any Single Vendor Irrelevant
Understanding this pattern was the same moment I started rebuilding my setup to outlive it. I didn't see the Gemini CLI shutdown coming by name, or predict Claude's safeguards, or guess OpenAI's pricing move.
Watch any group of vendors for more than a few months and the shape becomes obvious. Building around the shape costs almost nothing compared to getting blindsided by it. Three decisions carried most of the weight.
Decision one was routing. Rather than wiring my whole workflow to a single provider, I spread the work across Opencode Go, OpenRouter and Codex (via ChatGPT Pro) depending on the job.
One small file tells my agent which company handles which kind of request. Writing goes to one model because it holds tone better. Research goes to another because it chews through long documents faster. Routine generation goes wherever the price is lowest that hour.
When a provider hikes prices, throttles me, or pulls a model, I change three lines in that file and the work keeps flowing through the others without missing a beat. I broke down the full money side of this in my cost comparison against Claude Max, because the savings surprised even me.
Decision two was memory, and this one mattered most. Claude's conversations disappear when the chat closes. ChatGPT's history lives on OpenAI's servers, under OpenAI's rules. I know that both Claude and ChatGPT have memory, but it's a single file and you cannot alter it directly. All this means that you don't have full access to it.
My setup has 3 types of memory. First, Hermes comes with its own memory files and also built-in tools for session recall, so it can search our past convos. Then I added a 3rd party memory provider called Hindsight, but I might drop it as it's proven to be too much hassle for too little return for me.
Last but not least, I use Obsidian with plain markdown files, synced across my devices through the built-in Sync, readable by any AI tool I choose to point at it. Project notes, research, past sessions, saved agent routines, all of it sits as a file on my VPS, laptop and phone. I also configured a WebDAV server and MCP so literally any AI can get access to my files (behind a login of course).
A provider changes its deal tomorrow and my context follows me wherever I go next. None of it ever belonged to the provider in the first place.
Decision three was the agent layer, and this is the piece most people skip. I run Hermes as the stack that sits between me and whichever company happens to be doing the actual thinking.
Hermes holds the workflow, the saved skills, the routing, and the handoff logic. Whatever provider is plugged in becomes a swappable engine. Switching it means changing a single reference.
When Claude started refusing harmless prompts last week, my workflow never stopped. I would've sent the work somewhere else and kept moving. Every saved skill that taught my agent how my projects run stayed identical, because those skills live in files I control.
I moved to Hermes specifically for this, because the layer running my work needed to stay independent.
None of these three decisions asks for a computer science degree. Each one asks for a single choice, which is refusing to build your livelihood inside a tool that rewrites its own terms without asking you first.
How to Build Your Own AI System
Nobody builds the whole thing in two days, and trying to is how people give up. You build it one decision at a time, letting the setup grow around the parts you genuinely use rather than the parts a tutorial told you to want.
Order matters, because each layer makes the next one easier. Here is the order I followed and why.
Start by owning your memory. That move pays off the same afternoon and keeps paying off every month after.
Pull your project notes out of whatever AI chat tool currently holds them and drop them into a notes app you control. Obsidian, Notion, plain text files in a folder on your desktop, any of them works.
The brand of the app matters far less than the fact that nothing inside it disappears when you close a tab. Every month you keep building context inside a chat window is a month of thinking that disappears the moment the company changes its mind.
With your memory off the company's servers, adding a second provider becomes the natural next step instead of a scary one. Pick one piece of your workflow and run it through a different service for a week.
If Claude writes for you, send a research task through OpenRouter and watch what happens. If ChatGPT does everything, hand one job to a rival model purely to feel the difference.
Replacing your favorite tool is not the goal here. Learning what switching costs, while nothing is on fire, is the goal. That way the day you need to switch under real pressure you already know what you have to do.
Remember that your workflow is really a handful of separate jobs pretending to be one tool.
Ideation runs on one model. Research runs on another. Notes and memory live in your own app. These pieces talk to each other through plain files and simple handoffs.
The tool you love this month might not be the tool you want next year. Spend a little time now making sure that gap costs you nothing when it arrives.
Do This in Twenty Minutes Today
Fastest possible start, and also the simplest. Open whichever AI chat tool you lean on most.
Find the one conversation holding your most valuable project context, the thread full of decisions you made, research you gathered, and ideas you would really need to save.
Drop this prompt into that conversation:
Take the key decisions, facts, open questions, and next steps from this conversation. Summarize them as a structured note with these sections: Project Overview, Key Decisions Made, Important Context and Research, Open Questions, Next Steps. Write it as clean markdown I can paste into a note-taking app. Do not include our chat back-and-forth. Only the useful context.
Copy what comes back. Paste it into any app you own, whether that's Obsidian, Notion, Google Docs, or a single text file on your desktop.
Label it with the project name and today's date, and close the loop.
You're done. You moved your first piece of context off a company's servers and into something nobody gets to switch off.
Your chat tool still works exactly as it did five minutes ago, no harm done, no bridges burned. What changed is that the best part of that conversation now survives even if the tool rewrites its terms tomorrow morning.
Ten conversations fit comfortably in twenty minutes. And once you're done you'll figure out that you've already taken the first step towards building your own system. When you decide to install Hermes, or an open-source agent, you'll have everything you need to get going.
Get Ready For When Things Break
This setup is not perfect, and pretending otherwise would waste your trust. Some months I pour more time into maintaining my routing file than I would have spent living inside one vendor and taking the lock-in.
Independence carries an upkeep cost.
Memory sync across devices stumbles now and then. Mostly when I edit the same note on two machines at once and create a conflict I have to fix by hand.
Routing across providers means keeping three pricing models straight in my head instead of one. When a company changes its rates I need to update my calculations. But usually prices go down instead of up with each open source model (not the same can be said about proprietary ones).
A handful of tools refuse to hand context to each other cleanly. That leaves me copying and pasting between them in a way a single sealed vendor would have smoothed over for me.
Honest version, this whole approach suits people who already got burned by vendor dependency at least once and decided they would rather spend an hour a month on upkeep than risk losing an entire year of work to a corporate decision they were never allowed to influence.
Draw Your Own Dependency Map
Give this two more minutes. Write down three things in your AI life you couldn't easily replace.
One model that handles your most important work (like Opus is for some). One app holding your project history and notes. One workflow that breaks if a single tool changes its terms next week.
Now run one quick test on each. Try to export it.
Find your AI conversations locked inside a chat window with no real export button, and you found your first weak point. The twenty-minute exercise above already started repairing it.
Find your project notes trapped inside a tool that won't let you download them, and you found your second weak point. You already know its name without checking.
Export cleanly from all three and you hold more independence than almost every AI user out there. Fail to export from even one and you found exactly where your next twenty minutes belongs.
My own automated morning workflow runs this same provider-agnostic pattern on a schedule. Every check reaches into files I own instead of interfaces I rent by the month.
Gemini CLI's shutdown is only the newest entry in a pattern with no plans to stop. Claude's hidden safeguards, OpenAI's pricing war, Google's license reversal, each one is the same story wearing a different month on the calendar.
Everything above is the short version, the field notes. Full treatment, including the actual routing files, the Obsidian Sync setup, and the provider comparison that pushed me toward OpenCode Go and OpenRouter, is going into the first Hermes 101 course.
I'm building it right now and it should be ready soon. That course is where the patient step-by-step version of all this will live for anyone who wants their hand held through it.
If the money angle is what grabbed you, I wrote a full breakdown of the $30 Hermes stack that goes toe to toe with Claude Max and asks for no subscription. If you want the origin story, my migration from OpenClaw to Hermes is the longer answer to why I needed an agent layer I controlled before any of the rest of this made sense.
