ThinkRail picked up 1,173 stars in a single day and topped GitHub's daily TypeScript Trending list on October 27, 2026, with 3,194 in total. It is not another IntelliJ plugin: it is a lightweight IDE from JetBrains built around a coding agent, pi, where every task lives in its own git worktree and the pull request is handled from the top bar.
What it is and how it works
The premise is simple: if you are going to run several agents at once, each one needs its own branch and directory, and you need to see what changed without losing your own checkout. A project is a git repo, and each task is a workspace cut as a git worktree on its own branch. Agents work in parallel; you merge the good branch and delete the rest. Finished workspaces are shelved automatically when their PR merges or closes, or after a quiet week.
The agent is not JetBrains' own: it is pi, running in-process. The architecture has three layers:
- Engine host (
packages/server): an HTTP + WebSocket server onBun.servewith onepisession per chat. - The wire (
packages/contracts): a typed, versioned contract between host and UI. - The client (
apps/web): React 19, Zustand and Tailwind v4.
pi owns models, skills, context compaction and cost; ThinkRail owns the workspace, the editor and the wire, and the README stresses that it "never assembles prompts behind your back". What it adds around the agent:
- A real IDE: Monaco tabs, a file tree with git change marks, terminals, and review comments anchored to lines that you can send to the chat as context.
- Several concurrent chats per workspace, each with its own model and a live token and cost readout, plus subagents and a shared plan whose steps can be committed one by one.
- Specs: a graph of
SPEC.mdfiles next to the code that the agent reads and maintains throughspec_*tools, so intent outlives the session. - Portable Agent Skills (the ones you already keep for other agents are read in place) and MCP servers from
pi'smcp.json, with an in-chat confirmation for calls the server does not mark read-only. - Once a GitHub PR is open it shows CI checks and the review decision, and lets you hand failing checks or a diverged branch to the agent.
Providers: sign in with JetBrains AI or use any provider configured in pi; ThinkRail has no accounts of its own. Licensed under Apache-2.0.
Installing it
There is a desktop app (macOS 14+, Windows x64, and Linux x64/ARM64 with glibc 2.38+, i.e. Ubuntu 24.04 or newer) and a CLI that opens the same app in your browser. The CLI installer downloads the binary and checks its SHA-256:
curl -fsSL https://raw.githubusercontent.com/JetBrains/thinkrail/main/install.sh | bash
thinkrail ~/code/my-repoThat is a curl | bash: if you would rather not pipe a script into your shell, the releases page ships the binaries with SHA256SUMS for a manual install. You need git on your PATH and an authenticated model provider. Linux binaries are unsigned, and the macOS CLI is signed but not notarized yet.
What I like and what I don't
What sells me is that it tackles the real pain of working with several agents: worktree isolation, diff review and the PR loop in one window. Treating SPEC.md files as the source of truth is also a good way to stop the next agent (or the next person) from starting from scratch. And the pace is high: 577 commits, the v0.2.0 stable release of October 18, 2026 added MCP support and a hunk-by-hunk Changes review tab, and the nightly channel is already at 0.3.0-nightly.6 from October 25.
What gives me pause: it is a JetBrains incubator project at 0.x, and the README itself lists "full extension support" as upcoming. Basic telemetry (usage events and error reports tied to a random installation ID) is always on; only the extra usage data is opt-in. And it ties you to pi: if your team has standardised on Claude Code or Codex, you cannot swap the agent inside ThinkRail.
ThinkRail vs Claude Squad
The best-known tool for running several agents in parallel on worktrees is Claude Squad, a terminal app.
| ThinkRail | Claude Squad | |
|---|---|---|
| Approach | Full IDE around one agent | Terminal session manager for agents |
| Agents | pi (any provider it supports) | Claude Code, Codex, Gemini, Aider and other CLIs |
| Isolation | git worktree per workspace | Isolated git workspace per task |
| Requirements | git, a model provider | tmux and gh |
| Interface | Desktop or browser, Monaco editor | TUI in a single terminal window |
| License | Apache-2.0 | AGPL-3.0 |
| Maturity | v0.2.0, JetBrains incubator | Published releases, CI on GitHub |
When to use which. If you live in the terminal and already have a favourite agent, Claude Squad gives you per-task isolation without switching tools. If you want the diff, review comments, specs and the PR on the same screen as the agent, and you are happy to adopt pi, ThinkRail goes further. For a team, the licence matters too: AGPL-3.0 versus Apache-2.0.