What's new in 3.14
Data observability, approved fixes, AI agents and MCP in the UI, CLI 1.0, Helm chart 2.0 — and what to change when you upgrade.
3.14 moves CHouse UI from watching clusters to watching the data itself — every table, every ingestion source — and adds a way to fix what it finds, safely. Every AI feature now runs on an agent you manage in the UI, MCP moved into the UI, and the CLI reached 1.0.
Highlights
Data observability — The new Data page replaces DataOps and answers is the data right, right now? for every table, without scanning them. Freshness and volume baselines are learned from metadata from day one, every table gets a trust state, and criticality comes from real reads. → How data is watched
Pipelines and lineage — Materialized and refreshable views, Kafka, RabbitMQ, NATS, S3Queue, AzureQueue, MaterializedPostgreSQL/MySQL, external tables, Distributed inserts, dictionaries, async inserts, writers and scheduled jobs share one status vocabulary, and hidden failures surface as incidents. Lineage is built from metadata and query_log, with no instrumentation. → Pipelines, Lineage
Root cause and approved fixes — Incidents walk from the data symptom down to the engine with the evidence for each step and the blast radius. Fixes come from a closed catalog, run under a separate credential and only after approval (two approvers for high-impact changes), are verified, and can be rolled back. Approve in the UI, Slack or the CLI. → Incidents, root cause & fixes
Schema change preflight — DDL that would break a view, dictionary, job, promise or saved query is stopped with the impact, unless someone with schema:override confirms. → Schema preflight
Context for people and agents — Table descriptions, owners, grain and canonical metrics, dbt import, Chouse AI drafts of the empty fields (built from metadata and summary numbers, never raw rows), and plain-language watchers that become promises. → Context
Monitoring: Performance, Capacity, Upgrades — Regressions against each query's own baseline next to the changes around them; disk forecasts, codec trials and cost; upgrade readiness against your workload with canary replay. → Performance, Capacity, Upgrades
Agents — Every MCP and token session with what it read and cost, budgets checked with EXPLAIN ESTIMATE before queries run, health notices in results, and a pause switch. → AI Governance
AI agents in the UI — Every AI feature — the chat, Optimize and Debug, the Doctor scan, the diagnoses, the DataOps and observability assistants — runs on an agent you can see and change in AI Governance › Assistant: its prompt, model, read-only tools, skills, subagents and limits. Test a change on your own connection before you save it, roll back any version, and keep receiving CHouse's improvements to the built-ins you haven't changed. → AI agents, Customize an AI feature
Chat about CHouse itself — The chat has an agent picker. CHouse Admin answers read-only questions about CHouse — who can do what, what failed overnight, what changed this week — through its API as you, so it never shows more than your screens would. Auto routes each question to it or to ClickHouse Data and combines the answers. → Chat agents
MCP in the UI — The MCP endpoint is at /mcp on the web port and managed in AI Governance › MCP: turn it on, set allowed origins and timeout, and switch each tool on or off. Agents only see tools that are on and that their token allows. → MCP, tool catalog
CLI 1.0 — Tables in a terminal and JSON when piped, a hidden-prompt login, profiles, connection names, retries, --debug, custom CA trust, and new health, lineage, incidents, remediation, mcp and agents commands. → CLI, command reference
Also — Cluster-aware scheduled-query destinations (Replicated* on every replica, or per-shard tables plus a Distributed table), a distribution promise check, Doctor reports as notebooks with fixes, and role cards grouped by the server's permission categories.
Upgrading from 3.13
Migrations run on start; nothing to run by hand. Check these:
| What changed | What to do |
|---|---|
MCP moved to /mcp on the web port; port 8752 and all MCP_* settings (and mcp.* YAML keys) are removed | Remove them. After the upgrade, turn MCP on in AI Governance › MCP (it starts off) and point agents at https://<your host>/mcp |
Helm chart 2.0: mcp.* values, the <release>-mcp Service and the MCP Ingress are removed | Remove mcp: from your values; nothing else changes |
Fleet collection always runs; FLEET_POLLER_ENABLED is ignored | Remove it. FLEET_POLL_INTERVAL_SECONDS still works; OBSERVE_FLEET_INTERVAL wins |
| DataOps became Data | Old /dataops/* links redirect |
15 new permissions (observe:*, remediation:*, agents:*, …) | Built-in roles get their defaults automatically. Add them to custom roles that should use the new pages — see the permission catalog |
| AI prompts, tools and skills moved from code into the metadata database as built-in agents | Nothing to do: they are installed on first start and behave as before. Give ai_agents:view / ai_agents:manage to custom roles that should see or edit them in AI Governance › Assistant (Admin and Super Admin have both) |
| AI skills now actually load — before 3.14 the built-in skills were never offered to the model, and Bedrock deployments had extra planning and file tools | Expect AI answers to change slightly (usually better grounded). Raise an agent's step budget if a feature starts stopping at its step limit |
| The collector reads more system tables | Open Admin › Connections › Edit › Check privileges for each connection and run the GRANTs it lists |
| CLI 1.0 changed some commands and output | Re-check scripts: JSON is the default when piped, unknown subcommands now fail, and a few commands that only pointed at the UI were removed — see the CLI changelog |
To use fixes, add a remediation credential to each connection; for Slack approvals, set the SLACK_* variables. See Set up fixes.
The full list of changes is in the changelog.