Agent runtimes
Use the actual agent, with its own CLI, tools, and configuration. Keep a common interface for creating machines and streaming runs.
Runtime adapters normalize execution, not every native feature or conversation format.
Loading page…
Building blocks
A foundation for building your own agent harness: runtime and provider adapters, editable instructions, tools, and a remote workspace. Configure them in the dashboard or extend the MIT-licensed source.
Use the actual agent, with its own CLI, tools, and configuration. Keep a common interface for creating machines and streaming runs.
Runtime adapters normalize execution, not every native feature or conversation format.
Choose where your agent runs. Provider adapters handle provisioning, files, commands, and supported lifecycle operations.
Bring your own account. Compute is billed by the provider; sleep, snapshots, and ports differ.
Connect native model keys or supported routing services. Choose a model path that matches your runtime.
Hosted Claude Code uses Anthropic; hosted Codex uses OpenAI. Direct SDK compatibility differs.
Bring your own SKILL.md instructions, browse the package catalog, and configure the MCP connections your setup needs.
Catalog entries are not preinstalled or authenticated. Installation and runtime support are separate checks.
Edit the documents that guide your agent. Import existing instructions, export Markdown, and install documents on a selected machine.
Document export is not a complete harness backup. It excludes secrets, tool configuration, and full skill bodies.
Open the native CLI in your browser. Move between machines, inspect files and logs, or run commands from the SDK.
A terminal connects to real remote compute. Native PTY or tmux transport depends on the provider.
Configuration links open the dashboard and require an account. Inspecting or forking the source does not. Creating a machine uses your connected provider and model accounts.
Start from useful code, understand how it works, and adapt it to your workflow. That is the design principle—not a claim that this ships as a shadcn registry, a React component package, or a one-command harness installer.
A starter saves runtime suggestions, instructions, and selected abilities. It does not launch compute or authenticate tools. Connect your accounts, review the setup, launch a machine, and check the result in its terminal and files.
Read the setup guideToday you can fork the code, reuse SDK configuration, and export memory documents as Markdown. Full harness import/export, public publishing, and a setup marketplace are not available yet.