Use the AI Concierge
Catalog-grounded chat assistant. Floating button on every screen. Answers questions about your projects, on-call, runbooks, deployments, and SLOs in plain English.
The AI Concierge is a chat assistant grounded in your live MarBoba catalog. It answers questions about your projects, on-call, runbooks, deployments, and SLOs without you having to remember which screen has the data.
Two ways to open it
- Floating “Ask MarBoba” button — bottom-right of every signed- in screen. One click slides the chat panel in from the right.
- Sidebar → AI Assistant — opens the full-screen version.
Both share the same backend, so your conversation context is identical — pick whichever fits the moment.
What it knows
The Concierge classifies your question into one of seven intents and pulls relevant catalog data automatically:
| If you ask about… | …it pulls from |
|---|---|
| Vault secrets / expiry / rotation | Vault items + expiry timestamps |
| Deployments / pipelines / builds | Latest 10 deployments per app |
| On-call / who’s paging | On-call bindings + current rotation |
| Projects / apps / catalog / services | Project list + owners + lifecycle |
| Runbooks / playbooks / incident response | Runbook docs + linked apps |
| App health / scores | Scorecard snapshots |
| Pull requests | Pointer to the PRs screen (no AI needed) |
For anything off-script, the Concierge falls back to a general answer without grounding context.
Example queries
- “Who is on-call for billing?”
- “Which apps have secrets expiring this month?”
- “Show me the runbook for the stripe webhook outage”
- “What deployments failed today?”
- “List my projects with overall health below 70”
Triggering an agent action
When you describe an action that maps to a registered agent action — for example “rotate the stripe vault secret” or “rollback to last green” — the assistant’s reply includes a Propose action button. Clicking it opens the agent-runs screen with the action pre-selected. The Step 25 two-person approval flow still applies before anything runs in production.
Recognized phrasings:
- “rotate” + “secret” or “vault” → Rotate vault secret
- “scale” + “service” or “replicas” → Scale service
- “rollback” + “last green” or “previous” → Rollback to last green
Privacy and quotas
- The Concierge runs on Anthropic Claude Haiku 4.5 by default (cheap UI-helper bucket per the routing policy). Enterprise BYO-key orgs can override to a different provider on their org config.
- Each turn counts toward your AI Calls quota. Free orgs get 100 turns / seat / month; Team / Business / Enterprise plans scale up. Buy AI Boost packs to extend.
- The Concierge does NOT have access to vault secret values — only metadata (names, expiry, environment). Treat it like any other catalog reader.
- Conversations are not persisted today. Closing the panel clears history.
When the Concierge can’t help
- Anything outside the catalog (“what time is it in Tokyo”)
- Anything that needs write access. Use the relevant screen + the agent-action approval flow instead.
- Long historical analytics — those live in the Velocity Dashboard + DORA metrics screens.
See also
- Buy an AI Boost pack — $25/mo recurring add-on that extends your monthly AI quota by +1,000 AI Calls / +100 AI Generations / +100 Agentic Actions per active boost. Stackable on every paid tier.
- AI provider routing — How MarBoba picks the right AI model for each feature. Anthropic Claude is the default; OpenAI is available as a fallback.