train.sh

What It Is

TrainShell is a GPU Training Workbench from Grid Heap, Inc., positioned as a “terminal-native AI workbench.” Instead of turning the development workflow into a chat interface, it brings browser terminals, tmux sessions, host status, an encrypted vault, datasets, recipes, and outputs into a single workspace for connecting to cloud GPUs or private Linux hosts and running Codex, Claude Code, OpenCode, and long-running jobs.

Core Capabilities

Its main value lies in its BYO host fleet approach: you can connect Vast.ai, RunPod, private EC2 instances, Colab, Codespaces, or any Linux machine, without being forced to migrate to a separate hosted runtime. The workbench supports browser terminals, shared terminal viewing, host metrics, and connection status. Pro and above add a recipe workspace, sidecar editor, datasets/artifacts tracking, run state/result review, and reproducible experiment metadata. For secrets, the site emphasizes local browser unlocking and injecting credentials into tmux only after verifying host identity, which should appeal to users who care about credential management for remote training.

Pricing and Ecosystem

Starter is free and suitable for a first host and basic browser terminal use. Pro costs the equivalent of $7/month when billed annually, or $10/month when billed monthly, includes a 30-day trial, and unlocks the training workbench. Team is priced the same per seat and adds shared hosts, consolidated billing, onboarding, security review discussions, and more. Note that fees for GPU providers, cloud hosts, AI providers, and similar services are not included. Its ecosystem explicitly covers Vast.ai, RunPod, AWS EC2, Colab, Codespaces, Hugging Face, as well as Codex/Claude Code/OpenCode.

Pros and Cons

The advantages are that it is terminal-native, has a low learning curve, and suits engineers already comfortable with SSH/tmux/CLI workflows. It also avoids locking the training workflow into a single platform. The pricing structure is fairly straightforward as well. The drawbacks are that the collected text does not state whether it is open source, and there is no TrainShell self-hosting option. The API is only mentioned in the terms, with no concrete documentation details. Payment channels are described only as depending on the providers enabled for the account.

Who It’s For and Access from China

TrainShell is best suited to individual AI engineers, small teams, and users who frequently connect to cloud GPUs or private machines for training and Agent-based development. The main text does not provide details on access from China, and because it depends on external services such as RunPod, Vast, Colab, Codespaces, and AI providers, actual availability for networking, payments, and compliance needs to be verified case by case. Alternatives include VS Code Remote, native SSH/tmux, JupyterLab, Codespaces, or the built-in consoles offered by GPU platforms.