It's a mess
Almost every week, I ask Claude or ChatGPT to fix another flaw in the system I built to use them. I have multiple Claude and Codex subscriptions, run local models, and keep a self-hosted harness alive on a mini PC. Then there are the credits scattered across other services. For months, I have tried to turn all of it into one coherent setup that can move between providers. The models keep improving, and I keep playing whack-a-mole with my setup.
Every tool has its own island
Models are more than capable of driving my setup. The tools are the issue. They work, but each one works on its own island. Anthropic uses CLAUDE.md while other tools read AGENTS.md. Even tools from the same vendor behave differently. A skill I define for Claude Code in the terminal cannot be invoked in the same way from Claude Code in the desktop app. Project context from Claude Code does not simply carry into Cowork. Even naming them is a struggle. Claude Code CLI and Claude Code Desktop? They may come from different teams and serve different use cases. As a user, I still have to repeat the same instructions.
MCP standardizes how AI applications connect to external systems. In practice, an MCP server I connect to one Claude application is not automatically available in another. I still have to configure the same connection again.
The ground keeps moving
I struggled to keep track of agents running across several repositories, so I tried cmux, then moved to Emdash and started raving about it. A colleague pointed out that Claude’s desktop app had added similar functionality. A problem that had pushed me towards another tool a few weeks earlier could now be handled inside Claude. The tooling changes so fast that a setup can become outdated before I have fully adopted it. The subscription is the visible cost; continuously comparing, migrating and reconfiguring tools is the hidden one.
The laptop cannot be the system
I don't want my AI setup tied to a laptop. Vendors are adding remote-control features and we're even considering keeping our laptops always open while on the go. Tongue-in-cheek, but this can't be the way forward.
The answer is an independent layer for both individuals and companies. It runs continuously, knows your context and has access to your tools. Hermes already works this way for me. It runs on a mini PC, handles scheduled jobs, and can continue long-running tasks without relying on my laptop.
Unfortunately, it also proves the problem. Adding an independent harness means configuring, securing, updating and connecting one more system. The layer intended to unify everything becomes another piece of infrastructure to maintain.
Portability is the architecture
That makes portability more important than choosing the right tool today. A friend had built up memory and context inside a personal Claude subscription. When their employer gave them a paid work subscription, that context did not move with them. They had to extract it from one account and rebuild it in the other. The same problem applies to instructions, notes, ideas, skills and tool connections. If those live within a single provider, changing subscriptions or adopting a better model becomes a migration project. They need a source of truth outside any one AI product.
You have to decide which system owns your identity and context, which capabilities must remain provider-independent, and how much infrastructure you are willing to maintain to make that possible. Inside a company, it becomes an architecture decision: where context lives, who controls access, and what keeps working when a provider changes.
You either get locked into a vendor or you maintain an independent layer yourself.