Skip to content

AI

1 min read

Model features are off by default: nothing connects to a model and no AI control shows. The rules-based diagnosis of failures, plans and MCP work either way.

Turn it on

Section titled “Turn it on”
config:
  ai:
    enabled: true
    provider: anthropic        # or openai, for any OpenAI-compatible server
    model: claude-opus-5-5
    baseURL: ""                # for openai: such as http://vllm.example.com:8000/v1
    timeout: 60s
    features: {explain: true, draft: true, palette: true, admin: true}
    dailyTokensPerUser: 200000 # model tokens per person per UTC day
    maxInputRunes: 24000       # the most evidence sent in one call
    admin: {includeOwners: false}


ai:
  apiKeySecret:
    name: javo-ai
    key: api-key
Terminal window
kubectl -n javo-system create secret generic javo-ai --from-file=api-key=./api-key
  • provider: openai covers any OpenAI-compatible server: OpenAI, Azure OpenAI, vLLM, Ollama, llama.cpp. Set baseURL.
  • The API key comes only from the Secret named in ai.apiKeySecret, never from the values, and only the gateway gets it.
  • features turns each feature on or off: explain (Explain more on a failure), draft (Describe it on New app), palette (Ask AI in the search palette), admin (Ask about this cluster, for platform admins).
  • admin.includeOwners names app owners in what the admin question sends to the model.

MCP for coding agents

Section titled “MCP for coding agents”
config:
  mcp:
    enabled: true      # the default
    publicURL: https://javo.example.com   # for approve links; empty: https://<urls.uiHost>
  plans:
    ttl: 10m           # how long a proposed change waits for approval

Agents reach /mcp with a personal API token. They can read and propose changes; a person approves every change in the browser.