JetBrains Unveils ThinkRail: Agentic IDE Gains 1,211 GitHub Stars in a Day
JetBrains has released ThinkRail, a lightweight IDE in which an AI agent works not in your project folder but in a separate workspace on its own git branch. Several agents can edit code in parallel, and you keep the good results and delete the rest. The repository gained 1,211 stars in a single day. For anyone running coding agents, it offers ready-made scaffolding in place of homemade scripts built on git worktree.
AI-processed from GitHub Trending AI; edited by Hamidun News
Coding agents have quickly learned to write code, but organizing work around them is getting harder. An agent edits files directly in your working directory, two chats interfere with each other, and a week later nobody remembers why a module is structured the way it is. JetBrains has tried to solve this with ThinkRail, a new project that gained 1,211 stars on GitHub Trending in a single day.
ThinkRail is an open-source agentic IDE. It includes a Monaco-based editor, a file tree that highlights git changes, terminals, and a document viewer. The engine is the pi agent: ThinkRail runs it inside its own process and connects it to a web interface built on React 19. Models, skills, context compression, and cost tracking stay on the engine side, while the shell handles workspaces and the interface.
Workspaces Instead of a Shared Folder
The central idea is isolation. Any git repository can be opened as a project, and workspaces are created from it: each has its own branch and its own working directory. Agents work in them in parallel and do not touch the original source code. When a task is finished, you bring the needed results into the main branch and delete the rest.
You can simply write "do this in a side workspace" in the chat, and the agent will create a workspace and a new chat on its own, then track the progress of chats and pull requests. Finished workspaces close after a merge or a closed PR, or after a quiet week. When you return to work, the workspace comes back to life. The top panel shows how the branch relates to the main one and what state its PR is in. For GitHub, CI checks and reviewer decisions are displayed, and a PR can be merged or reopened directly from the interface.
There are also chats without a repository, for questions and drafts. Each has its own temporary folder. If a conversation turns into real work, the agent writes a brief context handoff and opens a new workspace chat with it.
Project Memory: Specifications and Skills
The second layer is long-term memory. Next to the code lives a specification graph, which the agent reads, searches, and updates through dedicated tools. Project intent survives across individual sessions. Development follows a spec-first principle: if an agent's change alters a module boundary or a contract, the specification is updated in the same change.
Reusable skills also work this way. ThinkRail can import them from Claude, Codex, and other coding agents. If the library is large, a compact list goes into the prompt, and the agent finds the rest by searching the skills. MCP servers are supported. Calls that a server does not mark as safe for reading require confirmation in the chat, and servers described in the repository are not enabled without your approval.
The interface is also designed for the agent. Instead of guessing, it asks structured questions with answer options, renders Mermaid diagrams, and shows before-and-after UI mockups that can be commented on like a diff. Mockups in the browser are clickable prototypes, and their scripts run in a sandbox with no network access. Outputs such as reports, diagrams, and screenshots go into a separate artifacts folder, and files deleted by the agent remain in the list.
Models, Platforms, and Privacy
ThinkRail has no account of its own. You can sign in through JetBrains AI or connect your own providers. Any OpenAI-compatible server is supported: LiteLLM, sglang, vLLM, Ollama, LM Studio, llama.cpp. This suits both a corporate gateway and a local model.
The application ships as a native desktop installer for macOS 14+, Windows, and Linux, as well as a CLI that opens the same interface in a browser. There are stable and nightly channels, and the installer verifies SHA-256 checksums. The interface is translated into seven languages, including Russian.
What to Consider Before Installing
There are caveats. The macOS CLI is signed but not yet notarized, and Linux artifacts are not signed at all. Older DMGs may be blocked by Gatekeeper. Versions are still at zero (the examples use 0.3.1 and 0.4.0-nightly), so the product is young. The application collects basic telemetry: launches, chat creation, message sending, and provider connections, tied to a random installation ID. Prompts, code, and files are not included. Error reports are always sent, while additional data can be disabled in settings.
The practical takeaway: if you already run several coding agents and assemble git worktrees, scripts, and notes for them by hand, ThinkRail is worth a look as a ready-made framework. Workspace isolation, the specification graph, and local model support address three common problems in this kind of work.
- Nov 21, 202622 Claude Agents Spent Six Weeks on Black Hole Problems — and Solved Two
- Nov 21, 2026Apple Is Building an AI Agent for Mac While Making It Harder for Outsiders to Get In: 14 Clicks Instead of 4
- Nov 21, 2026OpenAI Starts Rolling Out dots Background Agents to Some $20 ChatGPT Plus Subscribers
- Nov 21, 2026OpenAI to Give 50,000 Grants for Free ChatGPT Plus to 150,000 Scientists
Key Questions Answered
What are the key details on Workspaces Instead of a Shared Folder?
The central idea is isolation. Any git repository can be opened as a project, and workspaces are created from it: each has its own branch and its own working directory. Agents work in them in parallel and do not touch the original source code. When a task is finished, you bring the needed results into the main branch and delete the rest.
What are the key details on Project Memory: Specifications and Skills?
The second layer is long-term memory. Next to the code lives a specification graph, which the agent reads, searches, and updates through dedicated tools. Project intent survives across individual sessions. Development follows a spec-first principle: if an agent's change alters a module boundary or a contract, the specification is updated in the same change. Reusable skills also work this way. ThinkRail can import them from Claude, Codex, and other coding agents.
What are the key details on Models, Platforms, and Privacy?
ThinkRail has no account of its own. You can sign in through JetBrains AI or connect your own providers. Any OpenAI-compatible server is supported: LiteLLM, sglang, vLLM, Ollama, LM Studio, llama.cpp. This suits both a corporate gateway and a local model. The application ships as a native desktop installer for macOS 14+, Windows, and Linux, as well as a CLI that opens the same interface in a browser. There are stable and nightly channels, and the installer verifies SHA-256 checksums.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.