A real machine
behind every Slack thread.

Sparkyard gives every person in your workspace their own Linux computer and lets a coding agent work on it — Claude, Codex, Gemini or Copilot — from a thread, where the rest of the work already happens.

Not a container that vanishes. A microVM with your clones, your caches and your dotfiles still on it tomorrow — on hardware you own, billed to a model account you control.

How it reads

Mention it. That is the whole interface.

No dashboard to learn and no context to re-explain. The thread is the session: the agent works on your machine, asks when it is unsure, and hands back a link and a pull request — at the same time, because the machine is the same one.

What you get

Your computer, not a sandbox.

Every person gets their own microVM and disk. Clone a repo once and it stays cloned. Warm npm and Docker caches survive between sessions. Install a toolchain and it is still there next week.

Docker runs inside it. So do your tests, your build, and anything else you would run on a laptop — because for practical purposes it is one.

machine

4 vCPU · 8 GB memory, per person

disk

30 GB, persistent, yours

caches

npm, Docker and friends, kept warm on that disk

concurrency

3 threads working at once

idle

suspends after 15 minutes, resumes warm

Isolation

One tenant cannot reach another.

The boundary between workspaces is a virtual machine, not a namespace. Separate kernels, separate disks, separate network ports — and the ports are isolated at the bridge, so two tenants on one host cannot see each other's traffic at all.

Each VM proves it on every boot, and the machine does not accept work until it has. That is the trace above: it is not a mock-up of a check, it is the check.

Your models

Inference goes where you say.

Sparkyard does not have to be the model vendor's customer. Point a workspace at Bedrock, Vertex, Foundry, Azure OpenAI or a gateway of your own, and every turn is billed to your account and answered inside your cloud — the same way the code is already running on your hardware.

Set it once per agent, from Slack, with the credentials sealed rather than pasted. A half-configuration is refused instead of quietly ignored, and a turn that reached the vendor's own endpoint anyway is reported — a workspace that believes it is routed and is not has been telling its security team something untrue.

claude

Bedrock, Vertex, Foundry, or any Anthropic-shaped gateway

chatgpt

Azure OpenAI, or any OpenAI-shaped endpoint

gemini

Vertex AI, in your own GCP project

copilot

GitHub operates it — there is no customer endpoint, and we say so

Everything it does

The features, in one place.

Shared desks

For the people who do not write the code.

Bind a channel to a desk and it answers everyone in it — from its own machine, with its own GitHub identity, scoped to exactly what it needs. Nobody has to connect an account, and nobody borrows a colleague's.