Chat agents
CHouse Assistant: one agent with toolkits, UI tools and read-only investigators — the agent picker, upgrading from CHouse Admin, and your own chat agents.
- In the app
- Chobi › Chat › Chat agent
- Needs
- ai:chat
The chat in Chobi talks to CHouse Assistant by default: one agent that answers, investigates with read-only helpers, and makes changes for you — each checked by its tier. Administrators can add their own chat agents in AI Governance › Assistant.
CHouse Assistant
One main agent owns the conversation, your page, its memory and every write:
- Core tools, always there: your schema, read-only SQL, charts, who you are.
- Toolkits of write tools: Data & Query and the toolkit of the page you chat from are loaded up front; the assistant loads others (Scheduling, Data Health, Observe & Ops, Dashboards & Context, Access, Platform & Audit) with
load_toolkitwhen your question needs them. Only tools you hold a permission for appear, so the model chooses among a few dozen tools, not all 74. - Your screen.
navigateandopen_entityopen a CHouse page or item in your browser;fill_formputs SQL in a new editor tab for you to review. Nothing is saved or run by them, and they only ever open CHouse's own pages. - Plans. For two or more changes it lists the plan first, then asks for them together, so you confirm once.
It knows which page you opened Chobi on (this dashboard, this incident) and uses that as context.
Investigators
For deep or broad read-only questions it delegates to an investigator with its own context. Investigators can't change anything, and what they return counts as data, never instructions:
| Investigator | Covers |
|---|---|
| ClickHouse Data | Databases, tables, columns, SQL, query performance, running and slow queries, server info and charts — on your active connection |
| Operations Analyst | Scheduled jobs and runs, Data Health promises and incidents, alerting, observability incidents with their fixes and notebooks, Doctor reports, the fleet, query regressions, capacity, upgrades, collectors, dashboards |
| Access Auditor | Users, roles and permissions, data access policies, your personal access tokens (never the secret), SSO, ClickHouse users and roles |
| Platform Auditor | Connections, AI models, CHouse's AI agents and which feature runs which, AI activity (usage, cost, data sent to providers), guardrails, external agents and MCP settings, the audit log |
Things to ask:
- Who can edit connections? — the Access Auditor finds the roles holding
connections:editand the users holding those roles. - What failed overnight, and pause the job that keeps failing. — the Operations Analyst finds it; the assistant pauses it, undoably.
- Keep an eye on shop.orders until tomorrow. — a watch.
- Create a dashboard of inserts per minute and open it.
An investigator only gets the tools for which you hold a permission. Ask about users without users:view and the answer says you lack access to that area — the agent doesn't try to work around it. Results are capped (100 rows) and secrets such as SSO client secrets, channel webhooks and API keys are redacted before the model sees them.
The agent picker
The Chat agent menu next to the model picker lists every chat agent you may use (it appears when there is more than one, on screens wider than a phone). The default is the agent bound to the Chat feature, CHouse Assistant unless an administrator bound another one.
The choice is saved on the thread: switching threads switches the agent, and a new thread starts with the agent you had picked. When a thread's agent is deleted, the thread answers with the default agent; when it is disabled or no longer available to you, the chat asks you to pick another one. Each answer is labelled with the agent that gave it, and in the activity panel every tool call made by an investigator shows via its path — for example via chouse-assistant › access-auditor.
Which agents appear
An agent is offered in the picker when all of these hold:
- it is enabled and has no task template (feature agents never appear)
- it is a router, sits directly under a router, is the Chat feature's bound agent, or is nobody's subagent — investigators under another agent stay behind it
- the user holds at least one of its required permissions, or it has none (admins see every agent)
Required permissions also prune a tree: a subagent the user may not use is left out of that user's run.
Upgrading from earlier releases
CHouse Assistant used to be a router over ClickHouse Data and CHouse Admin. On upgrade:
- the Chat feature moves to CHouse Assistant if it was still on the shipped default; a binding an administrator chose stays;
- CHouse Admin is retired: agents that delegated to it delegate to its three specialists, features bound to it move to CHouse Assistant, and the agent is removed — unless an administrator customized it, in which case it stays as it is;
- ClickHouse Data becomes a read-only investigator; its write tools moved to CHouse Assistant.
Adding your own
Any chat agent you create in AI Governance › Assistant shows up in the picker by the rules above. Typical uses:
- a team assistant — ClickHouse Data's tools with a prompt and a skill about your schemas and naming
- a restricted helper — a narrower tool set, visible only to roles with a given permission
- an assistant of your own with
load_toolkit— it then loads toolkits the way CHouse Assistant does
Build a chat agent walks through one end to end.
Limits
A chat run stops after 2 minutes unless the agent sets its own timeout (CHouse Assistant allows 4, since it waits for investigators), and after its step budget. Picking a model in the chat overrides the agent's model for the agent and its subagents that have no model of their own.