ThinkRail

• Running two Pi agents on separate worktree branches of the same repo, comparing their diffs, and merging the better one
• Keeping a project's goals and module decisions in SPEC.md files that the agent reads before each change and updates as it builds
• Reviewing what an agent changed in a turn, keeping or reverting individual hunks, and asking the agent about a specific hunk
• Moving from the Pi CLI to a GUI without losing configured providers, extensions, or chat history
• Taking a workspace from plan to GitHub pull request, then handing failing CI checks back to the agent

ThinkRail is the most direct way to get an IDE around Pi without giving up what makes Pi attractive. It does not reinvent the agent, so your providers, extensions, and history carry over. What it adds is structure around Pi's deliberately minimal core: worktree isolation by default, turn and hunk-level review, a plan, subagents, MCP management with approvals, and a spec graph that keeps decisions in the repo. The catch is scope and maturity. It is an experimental incubator project at version 0.2, it only drives Pi, and it does not add approval gates for the agent's own commands. Pi users and developers curious about spec-driven work should try it. Teams that want one surface for Claude Code, Codex, and other agents, or vendor-supported tooling, should look at JetBrains Air, T3 Code, or Orca instead.

About ThinkRail

ThinkRail is not a new agent. It embeds the open-source Pi coding agent as a library, runs it in-process, and puts an IDE around it: Monaco editor tabs, a files tree with git change marks, terminals, diffs, and several concurrent chats per workspace, each with its own model and a live token and cost readout. Pi still owns the models, system prompt, skills, extensions, compaction, and session state, and ThinkRail shares Pi's ~/.pi folder, so providers, extensions, and chat history carry over if you already use the Pi CLI. The workflow starts by opening a local git repository as a project and cutting workspaces from it. Each workspace is a git worktree with its own branch and working directory, so several agents can work in parallel while your checkout stays untouched, and you merge the branch you want and delete the rest. Working directly in the project folder is possible but an explicit choice. Two ideas set it apart from other agent GUIs. First, specs: a connected graph of SPEC.md files with frontmatter lives in the repo, and a bundled pi-spec-graph extension gives the agent spec_* tools to search and maintain it, so decisions outlive a chat. Second, review: since version 0.2.0 every agent turn ends with a summary of changed files and a Review turn button, and the branch diff offers Keep, Revert, and Ask agent about this hunk on each hunk. Bundled extensions add web research, subagents, a shared TODO plan, and inline diagrams and UI mockups, and ThinkRail runs the MCP servers from Pi's mcp.json with an in-chat confirmation before calls that may change data. Workspaces can push their branch and open or update a GitHub pull request through your own gh CLI, then show its checks and review decision. ThinkRail ships as a native desktop app for Apple Silicon Macs, Windows x64, and Linux (x64 and ARM64), or as a thinkrail CLI that opens the same app in a browser. It has no accounts of its own: you sign in with JetBrains AI or connect any provider through Pi's auth. The tradeoffs: it is a JetBrains incubator project, which JetBrains defines as experimental and provided as is; it only runs Pi, not Claude Code, Codex, or other harnesses; and a docs site is still marked coming soon.