Vibe code without losing control.

ThinkRail gives every AI agent a clear plan, a separate place to work, and a visible trail of progress. Run multiple agents at once while your project knowledge stays organized.

Explore on GitHub

// QUICK START

Run ThinkRail in your terminal

curl -fsSL https://raw.githubusercontent.com/JetBrains/thinkrail/main/install.sh | bash

Open your project

thinkrail ~/code/my-repo

Start working

Create isolated Git worktrees and start agent sessions.

JetBrains InnovationHubApache-2.0open source

// Agent session

this demo is interactive

Message the agent… (@ files · / commands · Enter to send)

// Foundational principles

Designed for high-trust development

We do not believe in fully-autonomous software engineering. We believe in automated agent orchestration backed by strict human gatekeeping.

See what your agents are doing

No black boxes. Every terminal prompt, source code lookup, and tool invocation is rendered in real-time. Know exactly what your agents are exploring.

Clarify requirements before building

Agents write a dry-run plan and architectural outline before writing code. Review the approach, adjust the constraints, and authorize with confidence.

Run multiple agents without the mess

Each task runs inside its own isolated Git worktree. Run five distinct features in parallel without dealing with lock files, dirty indexes, or branch switching.

Keep your project's memory

Maintain an active knowledge map of your codebase. When agents finish, they document how they solved the task, preserving reasoning for the next run.

// Capabilities

Complete engine overview

[ 01 ]

Spec-first workflow

Convert natural language directly into testable feature specifications.

[ 02 ]

Parallel workspaces

Isolated workspaces running safe sandboxed dependencies.

[ 03 ]

Live change stream

Interactive diff view tracking every modified source file instantly.

[ 04 ]

Interactive questions

Agents pause and prompt you when requirements lack clarity.

[ 05 ]

Persistent memory

Automatic documentation tracking for historical context.

[ 06 ]

Multi-model runtime

Select models dynamically per individual task complexity.

01 / Development strategy

Spec-first architecture.

Before touching code, your agent outlines requirements, dependencies, and expected outcomes in a centralized spec file. Once approved, the agent implements tests first to verify expected behavior, preventing runaway side effects.

$ cat specs/payment-integration.md

## Requirements
- Integrate Stripe Elements checkout flow
- Handle synchronous webhook signature verification
- Ensure robust token rotation logging

✓ Spec validation passed. Ready to spawn agent.

$ git worktree list

/Users/thinkrail/dev/app d4c55b6 [main]
/Users/thinkrail/dev/app-auth-fix 7b4aef9 [auth-jwt-fix]
/Users/thinkrail/dev/app-tests-refactor da4c55b [tests-refactor]

Running 3 environments in parallel. No index contamination.

02 / Isolation environment

Parallel workspaces.

No more switching branches back and forth. ThinkRail leverages local Git worktrees to spawn multiple sandboxed instances. Each agent writes to its own isolated directory, preserving your primary development tree completely untouched.