Overview
Kintsugi is an experimental fork of AgentsInTheCloud for working with coding agents on self-hosted infrastructure. Its web interface brings together agent conversations, workspace files, Git changes and tools, with embedded VS Code and draw.io editing and previews. Native AMD64 and ARM64 builds support both Intel/AMD and ARM servers.
The project is being split into smaller packages for workspace provisioning, agent integration and UI features. The package split and interface are still evolving.
How it works
A Bun server manages workspaces and their agent sessions, provisioning containers through Docker on a dedicated Linux host. Workspace modules supply agent types, tools and views; live updates reach the browser through cable channels and Turbo Stream HTML updates.
The application keeps settings and workspace state outside the agent containers. GitHub connections and secret values are managed separately from agent sandboxes. Workspace tools are published as a read-only mount, while file and Git operations use the workspace's own environment.
Deployment uses separate application and system images. Release sources can be configured, including private GHCR repositories; the installer supports pinned updates and rollback.
Features
Coding-agent workspaces
Launch workspaces with agents, files, Git changes and task views in one browser interface.
Modular implementation
Bun and TypeScript packages separate workspace lifecycle, agent integration, shared UI and individual workspace features.
VS Code and draw.io
Edit code in the embedded VS Code interface, and edit or preview draw.io diagrams without opening a separate application.
AMD64 and ARM64 support
Native builds for Intel/AMD and ARM64 Linux hosts.
Workspace tools
Agent helpers are mounted read-only into workspaces, with stale helper files removed when the tool set changes.
Deployment controls
Configurable Docker image sources, private GHCR authentication, pinned releases and installer rollback.
Optional management API
A separate, token-authenticated management API, disabled by default. Non-loopback listeners require TLS; management bindings stay outside agent workspaces.
Experimental fork
Requires a dedicated Linux server with Docker. The package split and UI are still evolving; the original MIT licence and third-party attributions are preserved.
Architecture
Posts
2026-10-10
We Need To Start Seeing Other Agents