Resources · Claude connector

Claude Snowflake connectorWhat Claude can do in your Snowflake account.

The Claude Snowflake connector exposes 2 tools: 1 reads, 0 write, 1 is undocumented. Cortex Search reads your data; what Cortex Agents does is not described tool by tool. Below: what Claude reaches, what your Snowflake admin has to set up first, and where it stops.

Verified Trustpilot reviews · AI, automation & growth agency

Overview

What changes once Claude can reach Snowflake

Most people in a company never write SQL, so the numbers in Snowflake stay with the data team. The connector lets you ask Claude a business question in plain words and get the answer from your own warehouse. Snowflake's pitch is exactly that: natural language in front of structured and unstructured data, for business users too.

Search documents, not only tables. Cortex Search runs an unstructured search over data your company has indexed in Snowflake, such as contracts, tickets or product sheets, and brings the passages back to the conversation.

Ask a governed business question. Cortex Agents hands your question to an agent your data team built in Snowflake, which then picks among the resources it was given.

Keep one set of rules. Access goes through Snowflake roles, so a sales manager and a finance analyst asking the same thing get answers shaped by their own permissions.

Here is what the directory does not tell you. The two names on the sheet are tool types. What they reach depends on the server your Snowflake admin creates, and nothing works until that setup exists. Nothing runs by itself either: no tool fires when a table refreshes. For scheduled loads or alerts, the Snowflake n8n integration and the Snowflake Make integration do a different job, on a trigger.

Vocabulary

Five words before you connect

The vocabulary around this connector, in one minute.

Connector
The link you set up once between Claude and your company's Snowflake account, so Claude can query it while it answers you.
Tool
One named capability the connector gives Claude. Here there are two, and Claude calls them on its own while it replies.
Authorization
Snowflake's own sign-in and consent screen, where you let Claude act with your Snowflake identity. You can withdraw it later.
Approval
The confirmation Claude waits for in the chat before an action that would alter an account, shown at the moment it matters.
MCP
The shared standard behind connectors: it is what lets an assistant like Claude talk to an outside service such as Snowflake.
Connect

Connect Snowflake to Claude in three steps

  1. 01

    Find Snowflake in Claude

    Open Claude's settings, go to Customize then Connectors, and search for Snowflake. On a Team or Enterprise workspace, an Owner or Primary Owner enables the connector before members can each sign in.

  2. 02

    Start the connection

    Click Connect on its row and sign in to Snowflake in the window Snowflake opens. If the link breaks later, Disconnect and connect again. That window belongs to Snowflake, not to Claude.

  3. 03

    Read the consent screen

    Check Snowflake's consent screen before you accept. It belongs to Snowflake, and it frames what the access covers. Snowflake notes that Claude asks there for all your roles, yet the session still runs under the default role set on your user.

Tools

The 2 tools, split by what we know

Snowflake gives Claude 2 tools: 1 that read your account, 0 that change something in it, and 1 no official source describes.

One tool is documented as reading. The other one has no sentence in any official source that says whether it reads or changes anything. Names stay exactly as Claude shows them.

  • 1 read
  • 1 not documented

What Claude reads (1)

1 tool

One tool that searches your indexed data and leaves it as it is.

Undocumented (1)

1 tool

Snowflake describes how this tool type passes a question along, but no source says what it may do to your data. We will not guess it from a name.

Cortex Agents

No official source documents whether this tool only reads. Snowflake explains that it passes your message to an agent built by your company, which then answers; what that agent may do depends on how it was configured.

Watch out
the general approval rule below is what applies.
Approvals

What Claude checks with you first

By default, Claude stops and asks before each action it takes on an account for someone. The prompt appears in the conversation when it matters.

On Team and Enterprise workspaces, owners decide whether a member may let some actions through without being asked every time, and they can limit what a connector does across the organization, for instance reading only. Nobody overrides that from a personal account. Claude also works with your rights and nothing more. On the Snowflake side, that means your default role: Snowflake notes that Claude does not pick a role at sign-in, so the session runs under the one set on your user.

Plans

Which plans include it

None of the 819 sheets in the official directory shows plan availability. The answer per connector is published nowhere, which is a real hole in the catalog.

Anthropic publishes a general rule: remote connectors are open to all users on Claude, Cowork, Claude Desktop and mobile. On Team and Enterprise, an Owner or Primary Owner enables a connector for the organization before members can each connect. Snowflake adds a requirement of its own, an admin setup on your Snowflake account, described above. For the live state on your account, the connector's sheet in the official directory is the place to look.

Limits

Where this connector stops

A connector is not an automation. Claude calls these tools while it answers you, so nothing happens when a table loads or a metric drops.

The two names on the sheet are tool types; the real list depends on what your Snowflake admin exposes, set in the server object your admin creates. Snowflake's own description also mentions running SQL queries from plain language, which the two names on the sheet do not show. The tool list is an observed floor, not a promise. The partner badge in the directory is not a security audit, and Anthropic says on every sheet that it neither picks a publisher's tools nor guarantees their behavior. For a connector where two official sources disagree, see the Claude gmail connector.

Need help

Need help connecting Snowflake to Claude?

A person reads every message.

FAQ

Questions about the Claude Snowflake connector

01What can Claude do with the Snowflake connector?
Claude can put questions to your Snowflake data in plain language through two tools. Cortex Search runs a search over unstructured content your company indexed, and Cortex Agents passes your question to an agent your data team built, which answers from the resources it was given. What either one reaches depends on the server your Snowflake admin created and on the rights of your default role, so two companies can get very different results.
02Can Claude change data in Snowflake?
No official source says so for the two tools on the directory sheet. Cortex Search is documented as a search, which only reads. For Cortex Agents, Snowflake describes how the question is passed along but not what the agent may do, so the page leaves it as undocumented. Snowflake's description also mentions running SQL queries from plain language, and the real tools depend on the server your admin creates; the directory sheet shows only these two names.
03Does Claude ask before it acts in Snowflake?
The general rule is yes: by default, Claude asks for confirmation before each action it takes on an account for someone, and the prompt shows up in the conversation. No source describes a confirmation specific to Cortex Search or Cortex Agents, so that default is what covers them. On Team and Enterprise workspaces, owners decide whether members may let some actions run without a prompt, and they can restrict the connector for the whole organization.
04Which plans is the Snowflake connector on?
No official source publishes plan availability connector by connector, and none of the 819 directory sheets shows it. Anthropic's general rule opens remote connectors to all users on Claude, Cowork, Claude Desktop and mobile, with an Owner enabling them first on Team and Enterprise. On the Snowflake side, an admin still has to create the server and the security integration. The directory sheet, opened while signed in, is the only place showing the current state for you.
05Can Claude see everything in our Snowflake account?
No. Claude works with your rights, and in Snowflake those come from your default role, since Claude does not pick a role at sign-in. Access to the server does not open its tools: an admin grants usage on each search service and agent. Snowflake asks admins to make that default role the one holding the needed privileges. On Claude's side, an owner of a Team or Enterprise workspace can also narrow the connector for everyone, and members cannot undo it.
06Why won't Claude connect to our Snowflake account?
Usually because a Snowflake setting blocks it. If your account uses network policies, an admin must allow Claude's outbound IP addresses. Each user also needs a default role and a default warehouse, and sessions fail to start when the warehouse is empty. Finally, the server object and the security integration must exist before anyone signs in, since Claude runs its sign-in against that integration. The managed server is not supported in government regions either.
07Should I use Claude or an automation tool with Snowflake?
They cover different needs. Claude answers questions in a conversation: you ask, it searches or consults an agent, and nothing happens once you stop. It does not react when a table refreshes or a metric crosses a threshold. When you need scheduled loads, alerts or data moving between systems on an event, that belongs to an automation tool working on a trigger. For ad hoc questions from people who do not write SQL, the connector fits.