External AI applications for Cozystack. Deploy production-ready AI infrastructure — model serving, API gateways, notebooks, visual pipelines, image generation, SRE assistance — into your Cozystack cluster through the dashboard.

Every app appears as a first-class entity in the Cozystack dashboard — deploy with a click, configure through a generated form, scale and tear down independently.

HolmesGPT is not tenant-deployable: it needs cluster-wide read RBAC, so it is installed once by the cluster admin as an opt-in platform component (holmesgpt.enabled: true) rather than offered in the tenant dashboard. See docs/holmesgpt.md.

Two more apps — n8n (workflow automation) and Open WebUI (chat UI) — live in a separate catalog, cozyllm-nonfree. Their upstream projects are distributed under licenses that are not OSI-approved open source (n8n's Sustainable Use License, the Open WebUI License), so they are kept out of this repository for licensing reasons. Install the non-free catalog alongside this one if you need those apps and your use complies with the vendors' terms.

Apply the bootstrap manifest to a Cozystack 1.4+ cluster:

kubectl apply -f https://raw.githubusercontent.com/aenix-io/cozyllm/main/init.yaml

This registers a FluxCD GitRepository and a platform HelmRelease in cozy-system. Flux pulls every minute; within ~2 minutes you'll see five new ApplicationDefinition resources in the dashboard.

For pinning to a stable release rather than tracking main, see docs/upgrades.md.

For the full install walkthrough including verification, troubleshooting and uninstall, see docs/install.md.

┌──────────────────────────────────────────────────────┐
│            Cozystack Dashboard (PaaS)                │
│         Click → fill form → deploy CR                │
└─────────────────────┬────────────────────────────────┘
                      │
        ┌─────────────▼──────────────┐
        │   ApplicationDefinitions   │
        │  registered by this repo   │
        └──┬──────┬──────────┬───────┘
           │      │          │
   ┌───────▼┐  ┌──▼──────┐  ┌▼─────────────────┐
   │ vLLM   │  │ LiteLLM │  │  Front-end apps  │
   │ (GPU)  │◄─┤ gateway │◄─┤  Langflow        │
   │        │  │ +Postgr │  │  JupyterHub      │
   └────────┘  └─────────┘  │  ComfyUI (GPU)   │
   ┌────────┐               └──────────────────┘
   │ComfyUI │
   │(GPU)   │
   └────────┘

• vLLM serves models on raw GPU hardware
• LiteLLM unifies one-or-more vLLM endpoints (plus public OpenAI/Anthropic/etc.) behind one API
• Front-end apps talk to LiteLLM as if it were OpenAI — single key, model registry, usage tracking
• ComfyUI is independent: no LLM dependency, just GPU + Stable Diffusion

• Install guide — detailed install + verification + uninstall
• Upgrade policy — semver, ref pinning, schema migrations
• Application reference — per-app deploy options + usage
• End-to-end examples — wiring apps together: AI API stack from zero, shared ML notebooks, flows as APIs

Alpha. Pinned to specific upstream chart versions. Renovate opens PRs for patch bumps automatically (auto-merge once CI is green); major bumps need human review.

The code in this repository is licensed under the Apache License 2.0. The applications it deploys are fetched from their upstream sources at deploy time and retain their own upstream licenses — see NOTICE.

To add a new app, the easiest path is the cozystack:external-app-create skill from the cozystack/ccp marketplace — it generates a fully spec-compliant chart skeleton, dependency wiring (Postgres / Redis / Qdrant via Pattern C), and platform registration in one command.

Run make generate inside any packages/apps/<name>/ after editing values.yaml to regenerate values.schema.json and README.md via cozyvalues-gen. Mirror the regenerated schema into the corresponding openAPISchema field in packages/core/platform/templates/cozyrds.yaml.