Docker Agentic Platform
ExperimentalThe platform agents run on
The runtime layer for agent workloads
Start in minutes
Pick your agent. Choose a sandbox type, preconfigured and ready to launch.
Add your key once. It’s stored as a secret, never inside the sandbox.
Set the boundaries. Review access, tool, and compute before launch.
Run. Your agent works in the browser, even after you close the tab.
Configure. Connect. Run. Control.
Start with what you already run
Pick the sandbox type for the agent you already use, preconfigured and ready to launch. Add your model API key once; it’s stored as a secret, never inside the sandbox.
Add the tools you need
Authorize an MCP server once, from the curated catalog or by URL. Choose exactly which of its tools to expose, and every sandbox you launch can use them.
Give your agents their own machine
Your agent works in its own sandbox. It pulls your repos and ships PRs but never sees your keys; credentials are injected at runtime. Pause anytime and pick up where you left off.
Safety for every sandbox, every run
Network policies decide what a sandbox can reach. Use Docker-managed Open or Restricted policies, or write your own allow and deny rules, ready for every sandbox you launch.
Pay for Compute. Nothing else.
Sandbox compute
Billed per second while a sandbox runs, with the hourly rate shown before you launch. Pause a sandbox and billing stops with it.
Inference: bring your own API key
Anthropic, OpenAI, GitHub Copilot, or Google. Model calls runs on your own account and billed by your provider directly. We don’t mark them up or resell tokens.
|
Compute Size |
Best For |
Details |
Rate/ Hr |
|---|---|---|---|
|
Micro |
Quick scripts and small experiments |
1 vCPU / 2 GB |
$0.07 /hr |
|
Small |
Light coding tasks |
2 vCPU / 4 GB |
$0.14 /hr |
|
Medium |
Good for most tasks |
4 vCPU / 8 GB |
$0.28 /hr |
|
Large |
Bigger codebases and heavier workflows |
8 vCPU / 16 GB |
$0.56 /hr |
|
XL |
Builds, tests and big monorepos |
16 vCPU / 32 GB |
$1.12 /hr |
FAQs
What does ‘Experimental’ mean?
It’s not a beta and it’s not gated. Anyone can sign up and be billed for real usage.
What can I run?
Predefined sandbox types for Claude Code, Codex, Gemini CLI, Copilot, and OpenCode, each in an isolated, Docker-hosted sandbox with a live terminal.
How do my agents get to my repos?
Add a GitHub token once, then tell the agent which repos to clone. Credentials are applied through the sandbox proxy at request time; the value never lives in the sandbox, and the agent can still push branches and open pull requests. Public repos clone without any credential.
What does it cost?
Compute is billed per second while a sandbox runs, and you see the hourly price before you launch. Paused sandboxes don’t bill. Your first $25 of compute is included, and your model provider bills your own key for inference.
What happens when I pause a sandbox?
Compute billing stops and your work stays put. Resume it and pick up where you left off.
Can I run sandboxes on my own machine?
Yes, with Docker Sandboxes, which runs isolated sandboxes on your machine through the sbx CLI. Local and hosted sandboxes are separate for this release