Kintsugi is an opinionated fork of AgentsInTheCloud, originally created by Lucas Meijer. It runs coding agents on your own infrastructure, with your choice of model providers and tools for evaluating their work.
Features
Kintsugi combines reviewed upstream changes with fork-specific additions:
- Native AMD64 and ARM64 builds.
- A mobile-friendly, resizable composer.
- Built-in agent code mode, including classifier and image models.
- GitHub, GitLab.com and Bitbucket Cloud connections, plus a configurable Gitea HTTPS instance.
- A Changes view that includes unchanged tracked files, working-tree text editing, image previews and browser-supported audio/video playback.
- Dictation that preserves trailing audio and releases the microphone when recording ends.
- An embedded Draw.io editor and diagram previews.
- A scoped, token-authenticated management API.
- Configurable Docker release sources, private GHCR authentication, pinned updates and installer rollback.
- Word wrapping in the file viewer, including Markdown source mode.
- GitHub Copilot subscription usage, quota information and consistently labeled calendar-month pacing across Copilot and Codex.
- OpenCode Go five-hour and weekly usage tracking, with monthly exhaustion reflected in availability.
- Unattended installation and transactional admin bootstrap with private token delivery.
- Trusted LAN proxy access with configurable preview origins, Caddy TLS termination and persistent publications.
- A 1,000-origin preview pool (
41000–41999) with retained identities and pane-local failure handling. - Opt-in workspace LAN/private-network permissions, exact private TCP host/port grants, configurable CIDR groups and external DNS/mDNS forwarders, with restart status and security-scoped admin settings.
- Tag-only releases with native AMD64/ARM64 builds, staged checks, shared caches and selective CI.
The name comes from 金継ぎ, the Japanese art of repairing pottery with gold, because the original project was a brilliant way to glue together several components and I wanted to honor that approach.
You can read this blog post for the motivation behind this fork.
Installation
The current published release is v0.3.2. Releases publish app, System and workspace images and advance latest; stable is promoted separately. Publishing does not update or restart an installation.
Use a dedicated Linux server with Docker. The installer defaults to ghcr.io/rcarmo/kintsugi and ghcr.io/rcarmo/kintsugi-system; matching images must be available before installation.
curl -fsSL https://raw.githubusercontent.com/rcarmo/kintsugi/main/scripts/install.sh -o /tmp/install-kintsugi.sh bash /tmp/install-kintsugi.sh --action install
EROFS is the default. For a fresh installation on a host without EROFS support, use OverlayFS:
bash /tmp/install-kintsugi.sh --action install --storage-driver overlayfs
OverlayFS uses more disk space and initial provisioning I/O. Updates retain the selected driver; in-place switching is unsupported. See storage requirements for kernel and image compatibility.
On Linux with systemd, setup installs the private host updater by default. Settings uses Check now → Download update → Restart to update for System, the app and workspace image dependencies. Existing installations need one host-installer update from a release containing the service. --no-updater skips it; macOS and older installations retain app-only updates.
Setup asks for the app image repository and Tailscale machine name. System defaults to the app repository with -system appended. For locally built or separately named images, use --app-image and --system-image. Private GHCR sources accept --registry-config FILE. See automation for source selection, updates and rollback.
For unattended setup and LAN access, follow the deployment instructions. Use the installer from the selected release tag, with matching app and System version tags. Cloning the repository does not install a running service.
Upgrading an existing installation
Use Settings → Check now → Download update → Restart to update for a normal update. Saved release sources and explicit pins remain authoritative; select Latest if you want to follow new releases instead of a pinned version. Installation updates interrupt running workspaces, so finish active work first.
An app or normal installation update cannot rewrite an external Caddyfile. Installations still using the old ten-port preview pool need a one-time coordinated installer and Caddy update to expand their saved range to 41000–41999. Keep the existing origin assignments; origins are not recycled.
Generated Caddy configuration keeps TCP preview port 41641 available while disabling HTTP/3 on that listener, leaving UDP 41641 for Tailscale. Merge its listener options into your existing global Caddy block, preserve unrelated sites and settings, then validate and reload. An app update does not apply this change to an existing Caddyfile.
Stop active code-mode scripts before downgrading to an app that predates code mode. Image rollback preserves journals; it does not restore old session data. See release and downgrade guidance.
Git hosting
Open Settings → Git hosting to connect GitHub, GitLab.com, Bitbucket Cloud or Gitea. Repository search uses connected accounts; public GitHub search also works without a connection. Workspace images include gh, glab and bkt.
For Gitea, enter one server's HTTPS URL and personal access token. Custom ports and installation subpaths are supported, and the Git username comes from the validated account. The certificate must be trusted by the host. No additional Gitea CLI is installed. GitLab support currently targets gitlab.com, not arbitrary self-hosted instances.
Provider tokens stay outside Agent sandboxes. Workspaces receive placeholders that the egress proxy substitutes only for permitted destinations; Gitea credentials are restricted to the configured HTTPS origin, including its port. Connecting a private server does not grant workspace LAN access or automatically trust its SSH key. Configure network permissions and SSH trust separately.
Existing workspace containers do not gain new provider CLIs or environment variables merely because the app was updated. They need preparation against the updated workspace image; custom images must supply their own tools.
Management API
Kintsugi provides an optional token-authenticated JSON management API for host-local or LAN automation without Tailscale. It is disabled by default. Configure listeners and issue scoped tokens with apps/web/scripts/admin-bindings.ts.
Scopes cover workspace lifecycle, templates and shared configuration, secrets, SSH configuration, host diagnostics, and release/update management. Every request, including GET /openapi.json, requires Authorization: Bearer <token>. The security scope grants administrative power, including issuing tokens; grant it sparingly.
Non-loopback listeners require TLS, including listeners published through Docker. The installer publishes no admin ports unless explicitly requested with --admin-publish IP:HOST_PORT:ADMIN_PORT. Docker publications can bypass host firewall rules: bind to a trusted interface and enforce network filtering. See admin setup and permissions for configuration, certificates, token storage and revocation.
The normal app also supports JSON representations of UI operations for agents and Work views. Those use the app's existing access boundary and require a matching Origin on mutations; this CSRF check is not authentication. Agent prompting, arbitrary commands and terminal access are not exposed through admin listeners.
Development
bun install --frozen-lockfile bun run check bun run web
Internal packages use @kintsugi/*. Container accounts are agent:agent, with /home/agent as their home. The app uses UID/GID 1000; workspace startup aligns the numeric IDs with the host process to preserve shared-volume ownership. Existing ATELIER_* environment variables keep their names for compatibility.
Attribution and licence
Kintsugi derives from AgentsInTheCloud, originally created by Lucas Meijer. The original copyright and MIT licence are retained. Third-party attributions are in NOTICE.
Classifier-only provider credentials, OpenCode Go usage tracking and microphone lifecycle improvements were adapted from Larry Hudson’s fork. His authorship is retained in the imported commits.