Microsoft locks keys into the system vault
Microsoft has locked the keys into a system vault, so local models on the AI terminal finally don't have to "run naked." Microsoft has always claimed that running local models on the AI terminal (referring to Intelligent Terminal, the experimental terminal Microsoft forked from Windows Terminal) is most troublesome in "how to hand over the keys securely." The previous version only recognized key-free endpoints, so if you wanted to connect
💡 What You Will Learn
Microsoft has locked the keys into a system vault, so local models on the AI terminal finally don't have to "run naked." Microsoft has always claimed that running local models on the AI terminal (refe
📜 Table of Contents
Microsoft Locked the Key in the System Vault, and Local Models on AI Terminal Finally Don't Have to "Run Naked"
Microsoft has always claimed that the trickiest part of running local models on AI Terminal (referring to Intelligent Terminal, the experimental terminal Microsoft forked from Windows Terminal) is "how to hand over the key securely." The previous version only recognized key-free endpoints — if you wanted to connect a local model with an API key or an OpenAI-compatible service, there was simply no way in. Version 0.2.25 patches this hole.
What It Means for the Key to Go into Credential Manager (Windows' System Vault)
In the past, when many tools connected to local LLMs, the API key was just written in plaintext in a config file. The moment the config folder syncs to cloud storage, gets sent to a colleague, or gets picked up by some scanning tool, the key is running naked. This time Microsoft stores the key in Windows Credential Manager — the vault that comes with the system, the same protected place where your login password and WiFi keys live. The application only gets "permission to use this key," not the plaintext of the key itself.
This step looks minor, but it's actually a watershed: the security bar has been raised from "developers manage it themselves" to "the operating system manages it for you." For the models you run locally, the key no longer has to run naked.
In the official screenshot of the add screen, you can intuitively see what this flow looks like — Base URL, Model ID, and API key are clearly separated into three fields, with a special note that the key will be stored in Windows Credential Manager:
Why It's Worth Running Local Models in the Terminal for AI (My Honest Take)
I've always advised friends around me: if you have a GPU that can drive a 27B-parameter-class model, the value of running AI locally is very real — data never leaves the machine, it works offline, and there's no anxiety about per-token billing. Intelligent Terminal has been able to connect to key-free endpoints like Ollama since 0.2, and with this version adding key-based access, the combo of local models + cloud models is finally complete — one tab running Copilot in the cloud, another tab connected to your local open-source model, each doing its own thing without interfering with the other.
The Deeper Meaning of This Move: A Terminal That Sells "Privacy"
Look at it in the bigger picture: Microsoft is using "data stays on your own machine" as the differentiating selling point for AI-native applications. Automatic approval handles execution permissions, API keys in the vault handle access security, multi-agent tabs handle parallelism — the terminal, the most low-level tool of all, is turning into the "local base" for AI agents. Real engineering effort has gone into this; it's not something slapped together carelessly.
One Last Step to Go
Microsoft itself admits this is still experimental. Right now the scope of local model integration is still narrow, and there's room for improvement in the model ecosystem, performance monitoring, and multi-machine adaptation. But the direction is already clear: what Microsoft wants is for you to feel confident delegating authority to agents in the terminal, and "keys not running naked" is precisely the precondition for daring to take that step.
Will I set up a key for my local model? Once I get the 27B model running smoothly on this machine, most likely yes. Will you?
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.