JetBrains, maker of IntelliJ IDEA and PyCharm, has put an experimental AI-coding workbench on GitHub under an Apache 2.0 license. It is called ThinkRail, it is labeled a JetBrains incubator project, and it surfaced in developer news feeds around October 25, 2026. It is not a plugin for an existing JetBrains IDE. It is a separate desktop and CLI app built around an open-source coding agent named pi.
This explainer covers what ThinkRail does according to its own README, how it differs from the usual chat-in-an-editor pattern, and what Hacker News commenters said. Everything about features comes from the project page, and the criticism is clearly labeled as commenters' opinions.
Quick summary: ThinkRail is an Apache-2.0, JetBrains-incubator agentic IDE built around the pi coding agent. Every task can get its own git worktree, a graph of SPEC.md files holds project knowledge, and MCP tools that are not marked read-only need your confirmation. The repo showed about 7.3k stars and 656 commits on main when checked; a Hacker News thread had 242 points and 174 comments and was mostly skeptical.
What ThinkRail is
The README calls it "the agentic IDE that gets better every time you use it". It provides a workbench, isolated workspaces and durable project knowledge through specs and skills, and its stated next step is full extension support. The stack, per the Hacker News description, is Monaco for the editor with Bun, Electrobun and React. A daily digest on October 26 described the project as running the pi agent in-process and having about 2k stars at that point, so interest grew quickly afterward. (Sources: GitHub, Hacker News, Bitdoze)
The three ideas that matter
First, isolation. Projects can create workspaces as git worktrees, each with its own branch and directory, or attach existing worktrees. Finished or dormant workspaces move to a per-project shelf when their PR merges or closes, or after a quiet week, and return when you work in them again. That means several agent tasks can run in parallel without stepping on one another's files. (Source: GitHub)
Second, specs as memory. A connected graph of SPEC.md files lives beside the code. The agent reads, searches and updates them through dedicated spec tools, and a read-only view shows the graph as a tree. The project itself is developed spec-first: a change that alters a boundary, contract or decision must update the relevant spec in the same change. (Source: GitHub)
Third, a permission gate for tools. Servers from your pi mcp.json can be used or added in Settings. ThinkRail asks for confirmation in chat before calling any tool a server does not mark as read-only, though you can permanently allow trusted tools per project. Servers defined by a repository stay off until you approve them. (Source: GitHub)
Feature snapshot
| Area | What the README says | Catch |
|---|---|---|
| Workspaces | Git worktrees per task, shelf for finished ones | You need git on your PATH |
| Chats | Multiple concurrent pi sessions per workspace, each with its own model and token and cost readout; steer a running turn, queue follow-ups, delegate to subagents | Built around pi, not other agents |
| Git and PRs | Branch status vs origin, publish, push, open PR; with a GitHub PR, shows CI checks and review decisions, and can merge | Uses your own gh CLI for PR shipping |
| Models | Sign in with JetBrains AI or any provider through pi; custom OpenAI-compatible servers such as Ollama, LM Studio, llama.cpp, vLLM | No specific model list on the page |
| Skills | Agent Skills from other coding agents are read in place | Extension support is still the planned next step |
| Platforms | macOS 14+, Windows x64, Linux x64/ARM64 (desktop Linux needs Ubuntu 24.04 or newer) | Some builds unsigned or not notarized |
What the README admits
- The macOS CLI is signed but not yet notarized, and Linux artifacts are unsigned.
- Usage and error analytics are sent under a random installation ID; basic usage events and error reports are always sent, while additional usage sharing is a Settings switch (offered at first run) that can also be turned off per run with --no-analytics or THINKRAIL_NO_ANALYTICS=1.
- The project page showed 225 open issues and 124 pull requests when checked, so this is an active, early-stage codebase.
What Hacker News thought
The thread, with 242 points and 174 comments, was skeptical. These are commenters' views, not established facts. (Source: Hacker News)
- The most upvoted thread argued that an AI-written README signals low care, and one commenter questioned how SPEC.md files could be ground truth when they read like model output. Others replied that judging code quality by a README is superficial.
- Several commenters were disappointed by JetBrains' AI direction, citing the shuttered Fleet editor, the Air agent tool now being a plugin, and opaque token pricing.
- Some noted gaps: weak worktree support in the traditional IDEs, a pi-only focus with no Claude or Codex integration, and the choice of Bun and TypeScript.
- One commenter said agents should run in containers or VMs rather than on the local machine, a fair point since worktrees isolate files, not processes.
- Long-time JetBrains users praised the company's navigation and refactoring and said they would still try it.
Should you try it?
ThinkRail is most interesting as a reference design: one worktree per task, a spec graph the agent maintains, and per-call approval for any tool that can change something. Even if you never install it, those three patterns work with any coding assistant. For people who just want to chat with a strong model about code, a full IDE may be more than they need.
Copy the pattern without the app: for each AI-assisted change, make a separate git worktree, keep a short SPEC.md describing what the component must do, and review the diff yourself before merging.
01What is JetBrains ThinkRail?
An open-source (Apache 2.0) agentic IDE from a JetBrains incubator project, built around the pi coding agent, with git worktree workspaces, SPEC.md project knowledge and MCP tool approvals.
02Is ThinkRail free?
The software is Apache-2.0 licensed. You still need a model provider; it supports JetBrains AI sign-in, providers via pi, or local OpenAI-compatible servers such as Ollama.
03Which platforms does it support?
macOS 14 or newer, Windows x64 and Linux x64 or ARM64. Desktop Linux requires Ubuntu 24.04 or newer, and a CLI installer is available.
04How does ThinkRail handle MCP tools?
It asks for confirmation before calling any tool not marked read-only, lets you allow trusted tools per project, and keeps repository-defined servers off until approved.
05Does it work with Claude or Codex?
The README describes pi as the agent and says Agent Skills from other coding agents are read in place. Hacker News commenters noted a lack of direct Claude or Codex support.
Whatever coding tool you settle on, the underlying model matters most. LumiChats gives you 40+ premium AI models side by side in one subscription, so you can paste the same code problem into several of them and see which answers best.
