JetBrains releases ThinkRail, an agentic IDE that learns projects through living specs

An agentic IDE from JetBrains

JetBrains has released ThinkRail, an agentic IDE built around the pi coding agent that keeps durable project knowledge in a graph of "living specs". The project hit the Hacker News front page on October 25, 2026, and its code is public on GitHub under the JetBrains organisation. The repository pitches it as a lightweight but genuine IDE for "vibe coding" — one that adapts to your way of working and, in the project's framing, improves with use.

Specs as project memory

The premise is that a coding agent needs context that outlives a chat session. According to the project's README, ThinkRail keeps a connected graph of SPEC.md files beside the code itself. The agent consults, searches and updates this graph through dedicated spec_* tools, so design intent and project decisions survive between sessions instead of evaporating when a conversation ends. A read-only Specs view renders the graph as a tree inside the workbench.

The agent also picks up portable Agent Skills — skill files you may already keep for other coding agents — directly from where they live, and the IDE bundles guided workflows for setting up a project, brainstorming a design and shipping a pull request.

Isolated workspaces and the GitHub flow

ThinkRail opens any git repository as a project and can carve workspaces out of it as git worktrees, each with its own branch and working directory. Multiple agents can then work in parallel without touching your checkout; you merge the branch you want and discard the rest. Workspaces that go quiet are shelved per project automatically — once their pull request merges or closes, or after a week without activity — and reappear the moment you open them again.

A topbar tracks the active branch against origin and its pull request, and can publish, push or open a PR in one click, with the description drawn from the plan, the commits, your PR template or an agent-written draft. Once a GitHub PR is open, the IDE surfaces CI checks and the review decision, and lets you flag it as ready, hand failing checks or a diverged branch to the agent, merge using whatever method the repository used last, or reopen it.

A full workbench around the agent

Rather than a chat wrapper, the README describes a splittable workbench built from Monaco editor tabs, a file tree that marks live git changes, terminals and documents, all tied to whichever workspace is active. You can pin review comments to specific files and diffs and push them into a chat as structured context, or open the worktree in your installed IDE.

A workspace can host several pi sessions at once, each bound to its own model with a live readout of token usage and cost. You can steer a turn while it runs, queue follow-ups, delegate to subagents, and share a plan that both you and the agent edit, with each completed step committed and reviewed on its own. When the agent proposes a UI change, it can show a before-and-after mockup in the chat, which you annotate the way you would a diff.

MCP servers are supported through pi's mcp.json, with presets available in settings. ThinkRail prompts for confirmation before invoking any tool a server has not flagged read-only, leaves servers declared by a repository disabled until you approve them, and handles OAuth sign-in from settings.

Platforms and distribution

ThinkRail ships as a native desktop app or as a self-contained CLI that opens the same application in a browser; both embed the same engine host, and app state lives under ~/.thinkrail. The macOS build targets Apple Silicon only — no Intel variant ships — alongside Windows x64 and Linux x64/ARM64 builds that depend on GTK 3, WebKitGTK 4.1 and related libraries. Releases carry SHA256 checksums on stable and nightly channels, and updates are applied only after an explicit install-and-restart.

You either sign in with JetBrains AI or plug in your own provider credentials through pi's authentication; ThinkRail itself has no account system. English and Simplified Chinese are built in, and other languages can be added by dropping a translated JSON file into a locales folder. On the roadmap is full extension support, so the IDE can build the tools it needs around your actual workflow.

Why it matters

Agentic coding tools keep hitting the same two walls: context that evaporates between sessions, and agents that can wreck your working tree. ThinkRail takes a clear position on both — the spec graph is its durable-memory answer, while worktree-scoped workspaces keep parallel agents from trampling your checkout or each other. The more striking signal is who is shipping it. JetBrains is one of the biggest names in IDEs, and it is launching an agentic-first product as a separate, lightweight tool rather than bolting agents onto IntelliJ. That says something about where it believes developer tooling is heading. The open questions are familiar ones for spec-driven agents: whether the specs stay honest as a codebase drifts, and whether enough developers adopt pi as their agent of choice to make the bet pay off.