Microsoft Forked Its Own Terminal to Keep AI Out of Yours

Microsoft forked its Windows Terminal into a separate Intelligent Terminal to keep AI experiments out of the stable command-line experience, shipping bug-fix-only releases for the mainline while the fork adds agent features like error detection and slash commands. The Intelligent Terminal, announced at Build in June, uses the open Agent Client Protocol to support multiple AI backends including GitHub Copilot CLI, Claude Code, and local models, without embedding a specific model or requiring a Copilot account.

Dev Tools https://sourcefeed.dev/c/dev-tools Article Microsoft Forked Its Own Terminal to Keep AI Out of Yours Windows Terminal stays deliberately boring while the new Intelligent Terminal fork carries the agent experiments. Priya Nair https://sourcefeed.dev/u/priya nair Windows Terminal https://github.com/microsoft/terminal is back on GitHub's trending page this week, which is a strange place for a seven-year-old, 104k-star repo whose latest release is a bug-fix drop. On July 16, Microsoft shipped Terminal 1.24.11911.0 and Preview 1.25.1912.0 — a tab-dragging crash fix, an early-initialization workaround for a deadlock with NVIDIA drivers, a memory-corruption patch in the DRCS soft-font code — and quietly let slip that 1.26 is delayed with no ETA. That's the whole changelog, and it's exactly the point. The interesting work on Windows Terminal isn't happening in Windows Terminal anymore. It moved next door, into a fork called Intelligent Terminal https://github.com/microsoft/intelligent-terminal that Microsoft announced at Build in June — and the decision to fork rather than integrate is the most instructive thing Microsoft has done with AI on Windows in two years. The terminal became infrastructure, so Microsoft stopped touching it Some history explains the caution. Windows Terminal started in 2019 as the escape hatch from conhost.exe , a console host so constrained by backward compatibility that Microsoft's own team couldn't modernize it without breaking decades of Win32 console applications. Both still live in the same repo — Terminal, Preview, conhost, and their shared rendering and parsing guts — which is why this one project effectively is the command-line experience on Windows. Since Windows 11 22H2 made Terminal the default console, every cmd , PowerShell, and WSL session on a modern Windows box flows through this codebase. That install base changes the risk calculus. A regression in a hobbyist terminal emulator annoys enthusiasts; a regression here breaks CI scripts, ssh sessions, and enterprise logon shells at Windows scale. So the mainline cadence has turned deliberately conservative: servicing releases that fix crashes and driver deadlocks, a major version that slips without drama. For a default system component, boring is a feature. Which left Microsoft with a problem, because 2026 Microsoft very much wants agents in everything. The fork where the experiments went Intelligent Terminal 0.1, announced June 2 at Build, is Microsoft's answer: an MIT-licensed fork that installs alongside your existing Terminal winget install Microsoft.IntelligentTerminal , keeps all the tabs/profiles/themes machinery, and adds a docked agent pane Ctrl+Shift+. , automatic error detection with suggested fixes, slash commands like /fix and /model , and a ? prefix in the command palette that dispatches a prompt to an agent in a background tab. The 0.1.1 update in mid-June extended error detection to bash and WSL and added Windows 10 support; releases have kept coming since, with v0.1.1841 landing July 10. Terminal Chat — the old Canary-channel experiment — is deprecated in its favor. The architectural choice worth noticing is the Agent Client Protocol https://agentclientprotocol.com , the open JSON-RPC-over-stdio standard that came out of Zed. Intelligent Terminal doesn't embed a model or hardwire a Copilot backend; it spawns whatever ACP-compatible agent CLI you configure as a subprocess and pipes shell context to it. GitHub Copilot CLI is the default it'll offer to winget-install one if nothing's found , but Claude Code, Codex, and Gemini work out of the box, and XDA-Developers demonstrated the escape valve in practice: about five minutes to wire up the open-source Pi agent via pi-acp and run the whole thing against a local Qwen model on llama.cpp. No account, no cloud, no Copilot. That's the same play Language Server Protocol ran a decade ago — turn a per-editor integration problem into a protocol, and let the ecosystem compete on either side of the pipe. Microsoft adopting ACP rather than inventing a proprietary equivalent is a genuine signal that agent-to-tool plumbing is standardizing, and it puts pressure on Warp and the other AI-first terminals whose assistant is the product rather than a swappable subprocess. Fork-first is the lesson of Recall, applied It's worth saying plainly: this is not how Microsoft usually ships AI. Copilot buttons have been bolted into Notepad, Paint, and the taskbar whether users wanted them or not, and the Recall episode showed what happens when an ambitious AI feature lands inside a trust-critical surface by default. A terminal is about the most trust-critical surface a developer has — it's where credentials get pasted and production gets touched. Forking keeps experimental agent code, and its attack surface, out of the binary that tens of millions of people didn't ask to change, while still giving the agent-curious something real to install. The trade-off is real too, and Microsoft isn't hiding it. Version 0.1 feels like a 0.1: early users report performance issues, the feature set is thin beyond error-fixing and the chat pane, and a fork inevitably drifts — every month Intelligent Terminal diverges, the eventual merge-back gets harder. The 1.26 delay on mainline, unexplained, reads like a team whose attention is split; that's speculation, but forks aren't free. What to actually do If you live in Windows Terminal today: update and move on. The July servicing releases fix real crashes, and nothing agentic is coming to your default terminal uninvited. If you're already running Claude Code or Copilot CLI in a Terminal pane anyway, Intelligent Terminal is worth a side-by-side install — the error-detection loop fail a command, get a fix offered in place, no browser round-trip is the concrete win over copy-pasting into a chat window. Teams that can't ship shell output to a cloud model should look at the Pi/ACP route with a local model; it's the rare Microsoft AI product where fully local is a supported first-class path, not a workaround. And if you're building agent tooling, the takeaway is bigger than either app: with Zed, Google, and now Microsoft's terminal team on ACP, the protocol is starting to look like the LSP of the agent era. Target it once, and the terminals come to you. The mainline repo trending on a bug-fix week says developers still watch this project like infrastructure — because it is. Microsoft keeping the experiments in a fork is the first time in a while its AI strategy has respected that. Sources & further reading - microsoft/terminal https://github.com/microsoft/terminal — github.com - microsoft/intelligent-terminal https://github.com/microsoft/intelligent-terminal — github.com - Announcing Intelligent Terminal 0.1 https://devblogs.microsoft.com/commandline/announcing-intelligent-terminal-version-0-1/ — devblogs.microsoft.com - Windows Terminal 1.24 and 1.25 Update Released: Bug Fixes and v1.26 Delay https://www.ntcompatible.com/story/windows-terminal-124-and-125-update-released-bug-fixes-and-v126-delay — ntcompatible.com - Microsoft's new Intelligent Terminal isn't locked to Copilot https://www.xda-developers.com/microsoft-new-intelligent-terminal-not-locked-running-local-llm/ — xda-developers.com Priya Nair https://sourcefeed.dev/u/priya nair · AI & Developer Experience Writer Priya covers AI frameworks, developer productivity tooling, and the startup ecosystem across South and Southeast Asia, bringing a researcher's rigour and a practitioner's empathy to every story. She is deeply sceptical of benchmarks and asks hard questions so her readers don't have to. Discussion 2 This is the right call. We did the exact same thing at my last gig with our logging infrastructure — kept the stable version boring as hell while a separate team played with ML-powered anomaly detection in a fork. The moment you start experimenting in the main line, you're no longer maintaining a tool; you're maintaining a product roadmap, and that's when your core users start getting surprised by breakage. Smart to fork rather than flag it off. that's fair, but did the ML fork ever actually graduate features back to main, or did it just become a perpetual experiment that nobody ran in production? because i see the strategy—keep stable/boring separate from experimental—but i'm curious if microsoft has a real plan to harden any of the agent stuff before it touches actual users' shells, or if this is just a gentler way to let it live in the corner indefinitely.