How they work together
Each system is useful on its own. Together they act as one layer between you and the model: every prompt passes through all of them, on your machine, and each one makes the others better.
From the moment you press send to the moment the work is done.
You
In Cursor, Claude Code or Codex, in the editor or the terminal. Nothing new to learn and nothing to configure.
The code the prompt touches, how the session is going, and how you have judged difficulty before. The same one-line request can be trivial in one file and risky in another.
The most capable model at close to the lowest price for this task, with a one-line reason where you are working. Or it leaves your model alone, when the saving is small or the call isn't clear.
When a long conversation is compacted, the goals, decisions and approaches that already failed carry over in a short note, so whichever model runs next picks up where you left off.
As the agent works, secrets stay out of what it reads, commands you can't undo are warned about or stopped, and a session going in circles is called out.
You, again
Every switch and every catch is shown with its reason. You can always overrule a choice and keep your own model, and that choice teaches Adaptive Intelligence for next time.
What two systems do together that neither could do alone.
Efficient Router + Context Manager
A router alone can save money and still cost you the session. Because the working memory goes to whichever model runs next, routing and long sessions stop working against each other.
Efficient Router + Adaptive Intelligence
A fixed router treats everyone the same. Each prompt you keep on your own model teaches it how you judge difficulty, so its next picks land closer to the ones you would make.
Efficient Router + Guardian
The same guardrails apply whichever model runs. Moving routine work to a lighter model saves tokens without lowering the bar on secrets or destructive commands.
Guardian + Context Manager
When the session watchdog sees the same fix failing again, it says so. The approaches that already failed are part of the working memory, so the next attempt doesn't repeat them.
Cursor, Claude Code and Codex, in VS Code and in the terminal, all go through the same layer with one license. Move between them and the same judgment comes with you.
Each system has its own page with the detail and the numbers: Efficient Router, Context Manager, Guardian and Adaptive Intelligence. How we measured models is on Benchmarks.
Your console shows what it did for you, day by day. The Guardian stays free for good, and uninstalling is one command.