TLDR:

Workspaces give coding agents instant GPU access (baremetal / VM B200s, MI300X)
Zero infrastructure management, just tell your agent to use wafer target workspace to run, evaluate, or benchmark code
Pay by the second, only when your agent is actually running commands
File sync, command execution, and cleanup built-in

The Problem

AI coding assistants are writing more and more code, including GPU kernels. But there's a fundamental disconnect: agents generally run on your local machine, and have trouble accessing compute cleanly.

So you end up with a painful manual workflow:

1. Agent writes kernel code

2. You provision a cloud GPU instance

3. Set up SSH, Docker, environment

4. Sync code to the remote machine

5. Run the kernel, capture output

6. Copy results back

7. Paste into chat

8. Pay for idle time while you iterate

This breaks the agent's autonomy. The whole point of a coding agent is that it can iterate independently—but without GPU access, every optimization loop requires you to be the manual bridge.

The Solution: Workspaces

Workspaces give your coding agent direct access to GPU compute.

One command:

bash

wafer target workspace create dev --gpu B200 --environment baremetal

Your agent now has a B200 GPU. No provisioning. No SSH setup. No Docker configuration.

From there, the agent can work on this B200 as it if was operating locally:

bash

# Sync code to the workspace
wafer target workspace sync dev ./my-kernel/

# Run commands remotely
wafer target workspace exec dev -- python train.py

# Pull results back
wafer target workspace pull dev /workspace/results ./results

# Or SSH in for interactive work
wafer target workspace ssh dev

When you're done:

bash

wafer target workspace delete dev

Pay Only for What You Use

Traditional GPU development means renting an instance and paying for every minute—even when you're just reading code, thinking, or waiting for your agent to respond.

With Workspaces, you pay by the second, only when commands are actually running.

Kernel development is bursty. You write code, run it for a few seconds, analyze results, iterate. Most of your time isn't spent executing—it's spent thinking and coding. Why pay for a GPU that's sitting idle?

Workspaces flip the model. Your agent spins up compute when it needs it, runs the command, and you only pay for those seconds of actual execution. No more paying for idle GPUs while you debug a memory access pattern.

Demo

Claude Opus 4.6 autonomously creates a workspace, writes a GEMM kernel, and iterates on optimizations. No human touched infrastructure.

What's Supported

GPU Types:

B200 — NVIDIA Blackwell (180GB HBM3e, CUDA)
MI300X — AMD Instinct (192GB HBM3, ROCm)

Built-in Features:

wafer target workspace sync — Push local files to workspace
wafer target workspace pull — Pull remote files to local
wafer target workspace exec — Run commands remotely
wafer target workspace ssh — Interactive shell (baremetal only)
wafer target workspace list — See all your workspaces
wafer target workspace delete — Clean up when done

Get Started

bash

# Install wafer-ai
uv tool install wafer-ai

# Authenticate
wafer login

# Install the skill so Claude Code knows how to use wafer
wafer skill install -t claude

# Create a workspace
wafer target workspace create dev --gpu B200 --environment baremetal

# You're ready
wafer target workspace exec dev -- nvidia-smi

We'd love your feedback

What workflows and hardware would you like to see supported? What's missing? What's broken?

Reach out at emilio@wafer.ai or find us on Twitter/X.