Can I ask Tares what happened during crash loops straight from the console?

Yes, Tares has Ask, an assistant built into the console where you type a question about your data and it investigates, showing each tool call as it works. Ask needs a model provider on the instance: either ANTHROPIC_API_KEY in the daemon’s environment or one added under Settings, Model providers. It runs on the instance’s default provider, and the chat tells you which one that is.

How it works

Ask runs as a chat loop on the Tares daemon, with up to 10 tool rounds per question. Its tools are thin calls to the daemon’s own read endpoints, so it sees what the read API exposes and nothing more.

When you ask about a crash loop, Ask can use these read tools on the data already in Tares:

• List sources and connectors, and describe any catalog object (schema, entities, freshness, samples).
• Profile a source’s fields and list entities.
• Read recent events.
• List projects.
• Use read to get a correlated timeline, across every source or only one project’s.

Each tool call appears in the chat as it happens, and the answer ends with the verdict Ask reaches. You can follow how it got there.

Ask also helps when the crash-loop data looks incomplete. The same read tools answer questions such as “is anything not ingesting?”, “what looks sparse?” and “why is this label missing?” from live health and field coverage.

Two proposal tools, propose_labels and propose_trigger, draft catalog changes as cards you apply or skip. A proposed trigger names its sources, and when you apply it from Ask it goes to the Default project. The proposals change nothing by themselves. Apply goes through the same validated management APIs the console forms use.

Ask uses the default model of the instance’s default provider. On Anthropic that is TARES_AGENT_MODEL, which is claude-sonnet-4-6 unless you set it. The same providers power Tares agents.

Set it up

1. Start Tares and open the console. On a local setup, this starts the daemon:

tares up

2. Give the instance a model provider. Either set ANTHROPIC_API_KEY in the daemon’s environment, or open Settings → Model providers in the console and add one there.
3. Make sure the data from the crashing service is flowing into Tares through one of your sources. Ask reads what the instance holds, so the crash-loop events need to be there for it to explain them.
4. Open Ask in the sidebar, or press ⌘K from anywhere in the console. The chat names the default provider it runs on.
5. Type your question. For example:

Which service restarted the most in the last hour, and what happened right before each restart?

6. Watch the tool calls as Ask works, then read its verdict. Follow up in the same chat to narrow to one service, one source or one time window.
7. If Ask drafts a label or trigger card that would help next time, choose Apply or skip.

The empty state of Ask also offers starter questions for exploring data, debugging a source and organizing the catalog.

What you get

• An answer about what happened, built from the events, entities and timelines on your instance, without wiring up an external client first.
• A visible trail: every tool call Ask makes is shown, so you can check the evidence behind the verdict.
• A fast way to check that the data is there. If a source looks off or a label is missing, Ask can tell you from live health and field coverage.
• Draft labels and triggers as cards, so a crash loop you just investigated can become something Tares watches for. Nothing changes until you press Apply.
• One place to start. Ask is available from the sidebar or ⌘K on any console page.

For how Ask fits into incident work more broadly, see AI SRE. The Ask page covers the console assistant itself.

When something else fits

• Ask lives inside the console and reads this one instance. If you want an agent in your own workflow, such as Claude Code, Cursor or your own service, connect it over MCP to get the same read surface plus subscriptions.
• Ask answers when you ask. If you want Tares to watch for a crash loop and tell you, use a trigger, and let Ask draft it for you with propose_trigger.
• If no model provider is set on the instance, Ask cannot run until you add one.

Related