Intelligent Terminal: an ACP agent pane grafted onto Windows Terminal
A fork of Windows Terminal with native agent integration, right in your command line.
At a glance
- What is it?
- Microsoft's experimental fork of Windows Terminal adds a native agent pane that reads your shell output and proposes commands you approve. Here is what the README documents, what it leaves open, and who should wait.
- Who is it for?
- Adopt Intelligent Terminal if you already live in Windows Terminal on Windows 10 2004 or later and want an ACP agent to see your shell output without copy-pasting, accepting that the agent pane is an approval gate rather than an autonomous shell. Do not adopt it if you need a terminal on macOS or Linux, or if you want an agent that executes commands on its own.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 1 day ago.
- What is it written in?
- Mainly C++, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The copy-paste tax that the agent pane is built to remove
The problem is not that terminals lack AI. It is that the AI sits in a browser tab while the error sits in the scrollback, and the transfer between them is manual. Intelligent Terminal's answer is to put the agent inside the window that already holds the output. The README describes the agent as having "context on your shell output, no copy-pasting needed", and the Get Started flow says you simply start asking questions in the agent pane. That framing sets the audience: people who run a shell on Windows all day, hit failures, and currently paste them into a chat window. It is not aimed at someone who wants a scripted, headless agent. The project is a fork of Windows Terminal, so tabs, profiles, themes, shells and keyboard shortcuts are inherited rather than rebuilt. The agent is the delta.
ACP as the integration seam, and what the pane actually does
The mechanism is Agent Client Protocol. Intelligent Terminal does not ship its own model; it speaks to any ACP-compatible agent CLI installed on the machine. On first launch it auto-detects several of them (the README names Copilot, Claude, Codex, Gemini and OpenCode) and falls back to GitHub Copilot CLI, installing it via WinGet if nothing is found. The pane is a client, the CLI is the server, and the terminal is the host. Two details matter more than the architecture diagram. First, the agent does not run commands in your shell without explicit approval. The README gives three options for a suggested command: run it, copy it into your shell yourself, or dismiss it. That is a deliberate boundary, and it means the product is an advisor with a keyboard shortcut, not an autonomous executor. Second, context flows from shell output into the pane, which is what makes error detection meaningful rather than decorative. Ctrl+Alt+. opens the pane with error context already attached, which only works because the terminal owns the buffer the error was printed to.
Install paths, and the execution policy that will bite first
Three distribution routes are documented. The Microsoft Store listing is marked recommended because it auto-upgrades. WinGet users run winget install --id Microsoft.IntelligentTerminal -e. Direct downloads come from the releases page as an App Installer package for x64, arm64 and x86. The floor is Windows 10 2004 (build 19041). One prerequisite is not optional: a supported agent CLI and a subscription, with GitHub Copilot as the default. The README flags a specific failure before you hit it. If PowerShell reports "running scripts is disabled on this system" or an UnauthorizedAccess error, your execution policy is blocking your profile and shell integration cannot initialize. The documented fix is Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned. If you skip that, the terminal installs and the agent pane opens, but the shell integration layer that feeds it context is not wired up. There is a companion document, doc/installing-dependencies.md, for other dependency errors.
Configuration keys worth knowing before you commit
Settings live under an Agent section, and the README enumerates the toggles. Pane placement takes Top, Bottom (default), Left or Right. Error detection lets the terminal notice command failures; error suggestions, a separate switch, sends those detected errors to the agent for fix suggestions. Agent session tracking (hooks) controls whether active agent sessions and their status appear in the session management UI, and the README is explicit that turning Settings > Agents > Sessions off does not remove existing hooks, which may keep reporting activity; the setting governs installation and updates. Token usage display shows or hides token usage and cost in the status bar. Model selection is layered: a global agent and model setting applies to both the pane and the command palette, any pane can override with /model on the fly, and changing the global setting overrides every pane. Bring-your-own-key providers are configured with a Base URL, Model ID and optional API key against an OpenAI-compatible Chat Completions endpoint, routed through GitHub Copilot or OpenCode. API keys go to Windows Credential Manager, and keyless local endpoints such as Ollama are supported. Agents can also be pinned per profile, including agents installed inside a WSL distro.
The Windows lock-in is the design, not a gap to be patched
This is a Windows Terminal fork. The Store listing, the WinGet package, the App Installer download, Windows Credential Manager for keys, and the 19041 build floor all point the same direction. If your team runs macOS or Linux workstations, there is no documented path here, and the ACP agent CLIs it targets are cross-platform while the host is not. A second limitation is subtler. The approval gate that makes the pane safe also makes it slower than a shell-integrated agent that just runs things. If your workflow depends on an agent executing a multi-step fix unattended, this is the wrong tool, and the README's own description of the three response options confirms the design intent. A third is maturity: the project calls itself experimental, versioned 0.2.x, with a 0.1 line that started in July 2026 and a 0.2 line in August. The README also truncates mid-sentence in the session tracking paragraph, which is a small reminder that the documentation is still moving.
How this differs from running an agent CLI in a normal terminal
The obvious alternative is the status quo: keep Windows Terminal, install an ACP-compatible CLI such as GitHub Copilot CLI, and run it in a tab. That works, and it costs nothing new to learn. The difference is where the context lives. A CLI agent in a plain tab sees only what you paste or what it can read from files; it has no handle on the terminal buffer. Intelligent Terminal's error detection and Ctrl+Alt+. error-context shortcut exist precisely because the host and the client are the same process. The trade is that you give up the upstream Windows Terminal release cadence and take this fork's instead, and you accept a Windows-only build. If your agent work is mostly file editing and long-running tasks rather than reacting to shell failures, the plain-tab approach loses very little.
Maintenance, licensing and what the MIT grant does not settle
The repository is MIT-licensed, which permits commercial and private use, modification and redistribution provided the copyright notice and permission notice are preserved. That is the standard grant, and it does not cover the agent CLIs you connect to, your subscription terms, or the model providers behind a BYOK endpoint; those are separate agreements and this review is not legal advice. On maintenance, the release history shows a fast cadence: v0.1.18 in July 2026, v0.2 in August, v0.2.23 later the same month, with the last push in September 2026. Fast cadence on an experimental fork cuts both ways. You get fixes quickly, and you also get churn in settings that the README documents as still evolving. Store and WinGet installs track that cadence automatically, which is convenient until a build changes behaviour under you. Pinning a specific release from the App Installer downloads is the only documented way to hold a version, and the README does not describe a rollback procedure.
Editorial conclusion
Adopt Intelligent Terminal if you already live in Windows Terminal on Windows 10 2004 or later and want an ACP agent to see your shell output without copy-pasting, accepting that the agent pane is an approval gate rather than an autonomous shell. Do not adopt it if you need a terminal on macOS or Linux, or if you want an agent that executes commands on its own. Before installing, confirm which ACP-compatible CLI you will authenticate against, that your PowerShell execution policy allows RemoteSigned for the CurrentUser scope, and whether Settings > Agents > Sessions is on if you expect the session management UI to track anything.
Community notes