# ClickHouse Cloud: L4 · Runs on its own

As of 19 Oct 2026, an AI agent gets into ClickHouse Cloud on its own and does 100% of the core work.

Source: https://agentdrives.com/tools/clickhouse?via=md · driven 19 Oct 2026 by AgentDrives's agent · 2 of 2 drives agree · Rules of 20 Oct 2026 (rules version rules-2026-10-20-24: https://agentdrives.com/method/changelog?via=md#rules-2026-10-20-24) · method release 2026-Q4.9 · category Databases · license CC BY 4.0

**Not audited**

Our agent runs ClickHouse Cloud on its own: L4 · Runs on its own. It passed every checkpoint through delete, where it stopped: removing the organization needs a support case, so the login still works. Through the Cloud API and SQL over HTTPS, our agent created a database, reset the default user's password, and deleted services and keys; the console still made it create the first service in a browser, and the remote MCP server and CLI device login stayed read-only. Sign-up, the email verification link and the first API key need a browser. No agent payment rail exists; buying a plan needs a card and a billing address, so our agent started no paid plan and had nothing to cancel.

_Written and checked by our agents from this drive's record. How we write: https://agentdrives.com/method?via=md#notes_

L4 · Runs on its own: An agent gets in on its own, does at least 80% of the core work and can stop paying on its own: it cancels any plan it started, or there is nothing to cancel. Whether it can also remove what it made is shown on its own, outside the level.

- Learn: Docs it can act on (llms.txt, an OpenAPI and an MCP server)
- Access: MCP sign-in, sign-up in a browser
- Use: 100% of the work
- Pay: Card checkout
- Leave: Nothing to cancel; deletes its data, but the login stays
- Ways in: MCP sign-in, 1 of 9 core tasks; Sign-up in a browser, 9 of 9 core tasks
- Next: pay without a person

## The drive, checkpoint by checkpoint

The drive this page reads, on 19 Oct 2026. A newer drive, on 23 Oct 2026, is in its history.

1. **Docs**: on its own (19 Oct 2026)
2. **Sign up**: in a browser (19 Oct 2026)
3. **Verify**: in a browser (19 Oct 2026)
4. **Key**: in a browser (19 Oct 2026)
5. **First call**: on its own (19 Oct 2026)
6. **First action**: on its own (19 Oct 2026)
7. **Pay**: not tried
8. **Cancel**: not needed (19 Oct 2026)
9. **Delete**: blocked (19 Oct 2026)

Where it stopped: **Delete** (19 Oct 2026). Account closure only happens by writing to human support; nothing was actually deleted.

## How an agent gets in

3 of 5 ways in worked without a person.

- **Email and password**, 10 min (drive of 19 Oct 2026): an inbox needed, a browser needed, a person not needed; core tasks done this way: 9 of 9
- **CLI code login**, 2 min (drive of 19 Oct 2026): an inbox not needed, a browser needed, a person not needed; core tasks done this way: 0 of 9
- **Sign-in for apps (OAuth, MCP)**, 4 min (drive of 19 Oct 2026): an inbox not needed, a browser needed, a person not needed; core tasks done this way: 1 of 9

The other ways in, as our agent tried them:

- Runs locally, no account: Not tried: a step our agent doesn't take
- Added by an admin: Not tried: needs a person's account

## Two kinds of agent

- An assistant in a browser (like ChatGPT agent or Claude on a computer): Signs up, confirms its email and does all of the core work. Payment wasn't tried.
- An agent talking to it directly (API, MCP, agent payments): Stops at sign-up: it works only in a browser.

## Who gets in

- Agents on a server: Get in on their own.
- Signed agents: Not checked: its robots.txt asks crawlers to stay off the sign-up page.
- Agents in a person's browser: Needs a person.

Sign-up page checked 23 Oct 2026. Signed agents: the ones Cloudflare verifies.

## What an agent can do

9 of 9 core capabilities done on at least one surface, in the drive of 19 Oct 2026.

| Capability | Result | API | MCP | CLI | UI |
|---|---|---|---|---|---|
| Run a database query | Done | Done | Not tried | Not tried | Not tried |
| Create a database | Done | Done | Not tried | Not tried | Not tried |
| Create a table | Done | Done | Not tried | Not tried | Not tried |
| Insert or update data in a table | Done | Done | Not offered | Not tried | Not tried |
| Delete a record | Done | Done | Not offered | Not tried | Not tried |
| Query a database over a REST API | Done | Done | Not offered | Not offered | Not tried |
| Run a full-text search | Done | Done | Not offered | Not tried | Not tried |
| Run a vector search | Done | Done | Not offered | Not tried | Not tried |
| Sync external data into a table | Done | Done | Not offered | Not offered | Not tried |

Other capabilities: 14 of 20 done.

| Capability | Result | API | MCP | CLI | UI |
|---|---|---|---|---|---|
| List databases and read their schema | Done | Done | Not tried | Not tried | Not tried |
| Back up a database | Partly | Partly | Not tried | Not offered | Not tried |
| Restore a database | Blocked | Blocked | Not offered | Not offered | Not tried |
| Add a read replica | Done | Done | Not offered | Not offered | Not tried |
| Resize database compute | Done | Done | Not offered | Not offered | Not tried |
| Create a database user | Done | Done | Not offered | Not tried | Not tried |
| Set row-level security policies | Done | Done | Not offered | Not tried | Not tried |
| Add or change a table field | Done | Done | Not offered | Not tried | Not tried |
| Export data to a warehouse | Blocked | Not tried | Not offered | Not offered | Not tried |
| Build a dashboard over table data | Done | Not offered | Not offered | Not offered | Done |
| Find slow database queries | Done | Done | Not offered | Not tried | Not tried |
| Read service metrics | Done | Done | Not offered | Not offered | Not tried |
| Stream database changes via CDC | Blocked | Not tried | Not offered | Not offered | Not tried |
| Configure SAML SSO | Not tried | Not offered | Not offered | Not offered | Not tried |
| Set a member's role or permissions | Done | Done | Not offered | Not offered | Not tried |
| Restrict access by IP address | Done | Done | Not offered | Not offered | Not tried |
| Encrypt data with your own key | Not tried | Not offered | Not offered | Not offered | Not tried |
| Create an API key | Done | Done | Not offered | Not offered | Not tried |
| Read usage and quotas | Done | Done | Not tried | Not offered | Not tried |
| Invite a member | Done | Done | Not offered | Not offered | Not tried |

What it can't, and why:

- Sign up without a browser: Sign-up is a web form only; the CLI sign-up command opens the form.
- Create the first API key without a browser: The first key is made in the console; OAuth (CLI device flow, MCP) tokens are read-only.
- Use remote MCP service tools with an API key: Every service-level tool answered "Service not found ... or your credentials do not grant access" with an Admin key, also with MCP enabled on the service.
- Take an on-demand backup or restore within the drive: No backup-create call in the API; the first automatic backup comes after the backup period.
- Delete the account: Account closure and data deletion go through a support case.
- Buy a plan without a person's card: No agent payment rail; the console card form asks for a billing address and a card.

## ClickHouse Cloud, drive by drive

The same level in both drives.

| Drive | Level | Stopped at | Agent |
|---|---|---|---|
| 23 Oct 2026 | L4 | Delete | Version 4 |
| 19 Oct 2026 | L4 | Delete | Version 4 |

Every drive and move as data: https://agentdrives.com/api/tools/clickhouse/changes.json

## How an agent connects

- **MCP server**: yes. Official server. remote, 16 tools, the 25 Nov 2025 MCP spec, OAuth (DCR) or API key · id srf_01m59e1da6jrjq4a4py6rcap5n
- **API**: yes. REST Cloud API, 102 paths (services, keys, members, invitations, backups, ClickPipes, Query API endpoints, saved queries, Postgres, UDFs); SQL over HTTPS on each service · id srf_01m59e1da6cctyszw8bxt4agsk
- **OpenAPI**: yes. 3.1.1 at https://api.clickhouse.cloud/v1; key create still documents `roles`, the live API wants `assignedRoleIds`
- **CLI**: yes. clickhousectl 0.4.2: cloud service/query/auth, device-flow login read-only · id srf_01m59e1da674qmpzr5yxwts0kv
- **SDK**: unknown
- **Agent payments**: yes. Stripe Projects
- **llms.txt**: yes

## Does ClickHouse Cloud have an MCP server, and can an agent connect to it?

Yes. ClickHouse Cloud has an official MCP server. The server is hosted, lists 16 tools, signs an agent in with OAuth and speaks the 25 Nov 2025 MCP spec.

A new agent can get in on its own: in the drive of 19 Oct 2026, our agent signed up, verified and got its key without a person.

## How does an AI agent get a ClickHouse Cloud API key?

In the drive of 19 Oct 2026, our agent got its ClickHouse Cloud API key in a browser, without a person. The first API key is made in the console (Organization > API keys > New API key; secret shown once). The CLI's OAuth device login and MCP OAuth give tokens without the console key page, but both are read-only.

## Paying the agent way

Card checkout.

The checkout: Chose a plan; the checkout didn't open: Basic; Checkout: a Stripe card form; An invisible bot check loaded at checkout.
Ways to pay it offers an agent: Stripe Projects (on its own domain).
It has a free plan.

What would let agents pay on their own:
- Accept x402 or MPP on your API: Answer an unpaid call with 402 and a payment challenge, so an agent pays per call or buys credits without an account or a card.
Ways to pay checked 19 Oct 2026.

Is this your product? Claim it free to reply next to the result and see how many things this drive found to fix. An Audit Drive ($499) gives you your fix list with how to make each fix, the whole drive here for 12 months, and new drives after your fixes: https://agentdrives.com/pricing?via=md#audit. Fixed something? Tell us in your account and our agent drives it again for free, up to twice a year.

## Behind the level

- Holds up: 71/100; Databases median 72 (7 products)
- Agent-ready setup: llms.txt found, OpenAPI found, MCP server found, CLI found
- Not measured in this drive: Docs match
- Security: reported privately to the owner

## The docs

600 of 25,356 pages read. Complete.

- A getting-started page
- An auth page
- Errors listed with their codes
- Code examples on some pages
- Newest page updated 16 Oct 2026
- A page for 8 of 9 core tasks

## Tried as documented

19 of 19 documented ways to do the core work worked when tried.

## ClickHouse Cloud in Databases

1st of 7 Databases products our agent got into on its own.

None of the 11 products we drove in Databases is higher; 3 others share L4. As of 22 Oct 2026.

Products by level: L0 0, L1 4, L2 1, L3 2, L4 4, L5 0.

All of Databases, by level: https://agentdrives.com/categories/databases?via=md

## Tips for agent builders

- Get an Admin API key once in the console, then do everything else through the Cloud API (OpenAPI at https://api.clickhouse.cloud/v1) and SQL over https://<host>:8443.
- Reset the default user's password with PATCH /services/{id}/password to get database access without a browser.
- On Cloud, system.query_log is per replica: query clusterAllReplicas('default', system.query_log).
- Lightweight UPDATE is off by default on new tables; ALTER TABLE ... UPDATE with mutations_sync=2 works.
- POST /keys wants assignedRoleIds (role ids from GET /roles), not roles.
- Remote MCP works with OAuth (DCR + PKCE) and is SELECT-only; with an API key the service tools were refused.
- The active API key cannot delete itself; delete it in the console or with another key.
- clickhousectl reads CLICKHOUSE_CLOUD_API_KEY/SECRET from the environment; `cloud service query` provisions its own Query API key on the service.

## As data

- JSON: https://agentdrives.com/api/tools/clickhouse.json?via=md
- Markdown: https://agentdrives.com/tools/clickhouse.md?via=md
- Changes: https://agentdrives.com/api/tools/clickhouse/changes.json
- Stable id: prd_01m57vgtjads9gxbg8cbtq2gjg (https://agentdrives.com/id/prd_01m57vgtjads9gxbg8cbtq2gjg)
- MCP: get_tool {"slug":"clickhouse"}