Intelligent Terminal 0.2 runs agents on a local model

Most "AI in your terminal" features are a chat box wired to somebody's cloud API. Intelligent Terminal 0.2 is more interesting than that: Microsoft shipped bring-your-own-model support, so you can point the agent pane at a model running on your own machine. For anyone who has told a client "we can't put that repo through a hosted model," this is the version worth reading.

What actually happened

Per the Windows Command Line blog, the 0.2 release adds four things that change how you'd actually use it:

• BYOM providers. In the Agent Pane settings, open the BYOM section, click Add New, and give it a Base URL and Model ID. Anything that speaks the OpenAI API works — Ollama and compatible servers included.
• Per-tab agents. /agent switches the agent in one tab without touching the others. Built-in options include GitHub Copilot, Codex, OpenCode and Claude.
• WSL-native execution. Agents run inside your active WSL distro, so they see real Linux paths and tools instead of guessing across the Windows boundary. Windows and WSL profiles can each name their preferred agent.
• Session plumbing. /sessions searches past conversations, /move repositions the pane, and the pane now shows token usage, animated tool states, prompt history and image paste.

Install is winget install Microsoft.IntelligentTerminal, the Microsoft Store, or the GitHub releases page.

Why a local coding agent matters for your business

Three practical reasons, in the order they'll come up.

It removes a whole class of "can we?" conversations. Client code under NDA, health or payment data in a fixture file, an air-gapped environment — a local endpoint means the question stops being about vendor terms and starts being about which model runs well on the hardware you have. That's a much easier question.

Per-tab agent selection is cost control disguised as ergonomics. Run a local 7B against the noisy loop — log triage, one-line fixes, "why did this command fail" — and reserve a frontier model for the tab where the hard work happens. We wrote about the same pattern when Claude Code made subagents cheap: the cost lever is routing, not the model you like best.

BYOM is a portability test you can run today. If your team can point a terminal agent at Ollama and still ship, your dependency on any single provider is a preference rather than a lock-in. If it falls apart, you've learned something specific about where the dependency actually lives — usually in prompts tuned to one model's behavior, not in the tool.

The honest caveat: local models are slower and weaker than what you rent, and a laptop running inference is a laptop not running your build. Treat this as a second lane, not a replacement. But a second lane you own is worth having before the day you need it.

Sources: Windows Command Line blog, microsoft/intelligent-terminal releases.