Snowflake

The Snowflake app exposes the Snowflake SQL API and selected Snowflake REST APIs to Agent Factory agents through MCP and to Builder workflows through DSUL instructions. It runs in the installing workspace’s tenant context and covers SQL execution, asynchronous results, databases, schemas and virtual warehouses. Authentication supports platform-managed or tenant-owned Snowflake OAuth with PKCE, key-pair JWT, programmatic access tokens and pre-minted access tokens.

SQL and results

Execute one SQL statement, monitor or cancel asynchronous work, and retrieve paginated results

Database catalog

List, inspect, create, alter, drop and undrop Snowflake databases and schemas

Virtual warehouses

Inspect and manage warehouses, including explicit resume and suspend operations

Who is this for?

This connector is used by three different roles. Jump to the section that matches yours — each one is self-contained.

Agent builder

You build agents in Agent Factory and want them to query or administer Snowflake. → Agent builder tab.

Platform admin

You configure a shared Snowflake OAuth client and publish the capability. → Platform admin setup accordion below.

Workspace builder

You write Builder automations that call Snowflake operations directly. → Workspace builder tab.

Prerequisites (Snowflake side)

A Snowflake account and its public HTTPS account URL, for example https://myorg-myaccount.snowflakecomputing.com.
A Snowflake user or OAuth integration whose role has the privileges required by the operations you intend to expose. Use a dedicated least-privilege role rather than an account-admin role.
One supported credential:
- A Snowflake OAuth custom-client integration with refresh tokens and PKCE.
- A key pair assigned to the Snowflake user.
- A programmatic access token.
- A valid pre-minted OAuth or External OAuth access token.
A virtual warehouse for SQL that needs compute. Database and schema metadata operations may not require an active warehouse.

Platform admin (Governance) — one-time platform setup

Goal: configure a shared OAuth client for one Snowflake account, then expose the connector as a reusable capability in AI Governance.

1. Configure the connector

Create a dedicated Snowflake role

Create a role containing only the database, schema, warehouse and SQL privileges that agents need. Grant it to the users who will authorize the connector. Snowflake OAuth does not allow the connector to exceed the selected role’s privileges.

Create a Snowflake OAuth security integration

In the target Snowflake account, create a confidential custom-client integration. Use the core workspace callback as its redirect URI:

<api-url>/workspaces/slug:snowflake/webhooks/oauthCallback

On production, <api-url> is https://api.studio.prisme.ai/v2.

CREATE SECURITY INTEGRATION PRISME_SNOWFLAKE_OAUTH
  TYPE = OAUTH
  ENABLED = TRUE
  OAUTH_CLIENT = CUSTOM
  OAUTH_CLIENT_TYPE = CONFIDENTIAL
  OAUTH_REDIRECT_URI = '<api-url>/workspaces/slug:snowflake/webhooks/oauthCallback'
  OAUTH_ISSUE_REFRESH_TOKENS = TRUE
  OAUTH_ENFORCE_PKCE = TRUE
  ALLOWED_ROLES_LIST = ('PRISME_AGENT_ROLE');

Retrieve the OAuth client credentials

Run the following command with a role allowed to inspect the integration. Keep the returned secret in a secrets manager.

SELECT SYSTEM$SHOW_OAUTH_CLIENT_SECRETS('PRISME_SNOWFLAKE_OAUTH');

The result contains oauth_client_id and two rotatable client secrets.

Enter the credentials through the configuration app

Open the central snowflake workspace and launch its Configuration app at <studio>/apps/snowflake (for example https://studio.prisme.ai/apps/snowflake). Enter the Snowflake account URL, client ID, client secret, default role and optional warehouse. The app writes the central secret; do not edit Studio Secrets manually.

Tell workspaces to use the central client

Each consuming workspace selects oauthCentral in its connector configuration. The central client is tied to the Snowflake account configured above; another account must use tenant-owned OAuth or a non-OAuth credential.

2. Declare the capability in AI Governance

Create the Snowflake capability

Open AI Governance > Capabilities and create a capability named Snowflake.

Point it at the MCP endpoint

Set the MCP server URL to the connector’s MCP Endpoint and set Scope to:

context_id,agent_id,user_id

agent_id lets the connector identify and authorize the calling agent. user_id keeps delegated OAuth sessions user-specific.

Make it available to agent builders

Publish the capability in the catalog. Catalog visibility follows the organization’s existing RBAC; tenant-context capabilities do not require a separate per-capability role grant.

Smoke-test

Enable the capability on an agent, authorize that agent in the connector configuration app, connect a Snowflake user and run a read-only query such as SELECT CURRENT_USER(), CURRENT_ROLE().

Declaring the capability makes Snowflake available; it does not authorize an agent. Each installed connector instance still applies its Authorized agents allowlist. There is no OAuth auth-config JSON to attach in Governance.

Agent builder (Agent Factory)
Workspace builder (DSUL)

Agent builder

Goal: let an Agent Factory agent query and manage Snowflake through five MCP tools.

A Workspace builder must first install and configure the Snowflake app. Platform-managed OAuth additionally requires the one-time setup in the Platform admin accordion.

The MCP endpoint runs in the installed app instance. It resolves credentials server-side, identifies the caller from the capability scope and checks the workspace’s agent allowlist.

Option A — Enable the shared capability from the catalog

Open your agent

Open the agent in Agent Factory and go to its capabilities.

Enable Snowflake

Select Snowflake from the capability catalog. The endpoint and context_id,agent_id,user_id scope are already configured.

Connect to Snowflake

With OAuth, connect on the first tool call or from the connector configuration app. Key-pair, programmatic-token and access-token modes do not need an interactive sign-in.

The shared capability uses credentials and an allowlist managed by the platform workspace. Use a workspace-owned instance when the agent needs an isolated Snowflake account, role or credential boundary.

Option B — Run it from your own workspace (recommended)

Prefer a workspace-owned instance for production. Its credentials, OAuth tokens and authorized-agent list are isolated to your workspace, and its Snowflake role can be restricted to the agent’s exact duties.

Install and configure Snowflake

Follow the Workspace builder tab. Select an authentication mode and test the credential.

Allowlist the agent

In the configuration app, open Authorized agents, select the agent and save. The Install capability action can perform this wiring.

Add the MCP capability

Point the capability at this app instance’s MCP Endpoint and set Scope to:

context_id,agent_id,user_id

Connect the user when using OAuth

Click Connect in the configuration app or follow the connection prompt surfaced by Agent Factory.

Brief the agent in its system prompt

You have access to the Snowflake MCP tools: sql, results, databases, schemas and warehouses.
Each tool requires an `action`. Prefer read operations unless the user explicitly requests a mutation.
Before executing DML, DDL, database/schema deletion, or warehouse resume/suspend, restate the target
and obtain confirmation. Narrow SQL results with SELECT columns, WHERE, aggregation and deterministic
keyset pagination; never assume an omitted result row can be recovered from MCP output.

Legacy AI Knowledge agents without the native capability picker can add an MCP tool under Advanced > Tools > MCP and paste the MCP Endpoint. The agent must still be allowlisted and its identity propagated to the connector.

Restricting to read-only (least privilege). Snowflake permissions are role-based. For OAuth, set ALLOWED_ROLES_LIST on the security integration and configure the connector to request a dedicated read-only role. For a programmatic access token, use ROLE_RESTRICTION; for key-pair or access-token modes, restrict the backing user’s grants. The connector remains usable for reads while Snowflake rejects writes with 403 or a SQL privilege error. This provider-side role boundary cannot be widened from the Prisme.ai workspace.

Available Tools

Each tool takes an action argument that selects the concrete operation.

Tool	Description
`sql`	Execute one SQL statement, inspect its status or cancel it. MCP responses are capped at 100 rows and 80 KB.
`results`	Fetch a page from an asynchronous Snowflake Resource API result by opaque `result_handler`.
`databases`	List, get, create, create-or-alter, undrop and delete databases.
`schemas`	List, get, create, create-or-alter, undrop and delete schemas within a database.
`warehouses`	List, get, create, create-or-alter, resume, suspend and delete virtual warehouses.

Output Formats

Snowflake tool calls return one bounded JSON payload in MCP text content. The connector:

limits SQL execution to one statement;
asks Snowflake for at most 100 rows and caps the final MCP response at 80 KB;
omits binary cell values;
adds _mcpLimits and, when truncation occurs, _truncation metadata with a refinement suggestion.

The schema accepts outputFormat (verbose, structured or both) for compatibility. The current Snowflake implementation returns the same bounded JSON text envelope for all three values.

Tool Details

sql

{
  "name": "sql",
  "arguments": {
    "action": "execute",
    "statement": "SELECT REGION, SUM(REVENUE) AS TOTAL FROM ANALYTICS.SALES GROUP BY REGION ORDER BY TOTAL DESC",
    "warehouse": "ANALYTICS_WH",
    "role": "PRISME_AGENT_ROLE",
    "async": false
  }
}

Parameter	Required	Description
`action`	Yes	`execute`, `status` or `cancel`.
`statement`	For `execute`	One SQL statement. It may be read-only or mutating.
`statementHandle`	For `status` / `cancel`	Handle returned by this connector’s `execute` action.
`database`, `schema`, `warehouse`, `role`	No	Case-sensitive execution context.
`bindings`	No	Snowflake SQL bind values.
`async`	No	Return a statement handle without waiting for completion.
`requestId`	No	UUID used to make submission idempotent.
`partition`	For additional SQL partitions	Zero-based provider partition requested by `status`. It cannot recover rows omitted by the MCP response cap.

results

{
  "name": "results",
  "arguments": {
    "action": "get",
    "result_handler": "opaque-result-handler",
    "page": 0
  }
}

Parameter	Required	Description
`action`	Yes	Only `get`.
`result_handler`	Yes	Opaque handle returned by an asynchronous Snowflake Resource API operation.
`page`	No	Zero-based result page; defaults to `0`.

databases

{
  "name": "databases",
  "arguments": {
    "action": "list",
    "like": "ANALYTICS_%",
    "showLimit": 25
  }
}

Parameter	Required	Description
`action`	Yes	`list`, `get`, `create`, `createOrAlter`, `undrop` or `delete`.
`name`	For get/createOrAlter/undrop/delete	Case-sensitive database identifier.
`body`	For create/createOrAlter	Complete Snowflake Database resource definition.
`createMode`	No	`errorIfExists`, `orReplace` or `ifNotExists`.
`like`, `startsWith`, `fromName`, `showLimit`, `history`	For list	Server-side list filters and pagination controls.
`restrict`	For delete	Keep `true` unless cascading removal is explicitly requested.

schemas

{
  "name": "schemas",
  "arguments": {
    "action": "create",
    "database": "ANALYTICS",
    "createMode": "ifNotExists",
    "body": {
      "name": "REPORTING",
      "comment": "Curated reporting layer"
    }
  }
}

Parameter	Required	Description
`action`	Yes	`list`, `get`, `create`, `createOrAlter`, `undrop` or `delete`.
`database`	Yes	Parent database identifier.
`name`	For get/createOrAlter/undrop/delete	Schema identifier.
`body`	For create/createOrAlter	Complete Snowflake Schema resource definition.
`kind`	No	`transient` or permanent (empty value).
`restrict`	For delete	Keep `true` unless cascading removal is explicitly requested.

warehouses

{
  "name": "warehouses",
  "arguments": {
    "action": "suspend",
    "name": "ANALYTICS_WH",
    "ifExists": true
  }
}

Parameter	Required	Description
`action`	Yes	`list`, `get`, `create`, `createOrAlter`, `resume`, `suspend` or `delete`.
`name`	For all but list/create	Warehouse identifier.
`body`	For create/createOrAlter	Complete Snowflake Warehouse resource definition.
`createMode`	No	`errorIfExists`, `orReplace` or `ifNotExists`.
`ifExists`	No	Make resume, suspend or delete succeed when the warehouse is absent.

Resuming a warehouse starts compute billing. Suspending it can interrupt workloads. Call either action only after explicit user confirmation.

Workspace builder

Goal: install Snowflake in a workspace, configure credentials and call its operations from automations.

Installation

Go to Apps in your workspace.
Search for Snowflake and install it.
Open the auto-populated Configuration app, select an authentication mode, test it and configure Authorized agents.

Configuration

Field	Description
Configuration app	Auto-populated on install. Opens authentication, connection testing and the authorized-agent allowlist.

Auth mode	What you provide	Best for
`oauthCentral`	Connect a user; optional role and warehouse	A platform-managed OAuth client tied to the configured Snowflake account
`oauth`	Account URL, OAuth client ID/secret, optional role and warehouse; then Connect	A workspace-owned Snowflake OAuth integration
`keyPair`	Account URL and identifier, username, public-key fingerprint, PEM private key, optional passphrase, role and warehouse	Non-interactive service users with key-pair authentication
`programmaticAccessToken`	Account URL, programmatic access token, optional role and warehouse	Rotatable non-interactive access with Snowflake role restriction
`accessToken`	Account URL, pre-minted token, optional role and warehouse	Caller-managed OAuth or External OAuth tokens

Credentials and the allowlist are stored as workspace secrets and resolved server-side. The public account URL must be an HTTPS Snowflake account origin, not a Snowsight page URL.

Available Instructions

Every instruction resolves credentials from the workspace configuration and returns the raw Snowflake API payload.

SQL and asynchronous results

Instruction	Description	Returns
`submitStatement`	Submit one or more statements in App mode; `statement` is required and execution context is optional.	A SQL API ResultSet (`data`, `resultSetMetaData`, `statementHandle`) or QueryStatus for asynchronous work.
`getStatementStatus`	Check execution and fetch a result partition by `statementHandle`; optional `partition` is zero-based.	A ResultSet when complete or QueryStatus while running.
`cancelStatement`	Cancel a running statement by `statementHandle`; optional `requestId` identifies the request.	Snowflake cancellation status.
`fetchResult`	Fetch page `page` from a Resource API asynchronous `result_handler`.	The raw Result API page and its provider pagination metadata.

Databases

Instruction	Description	Returns
`listDatabases`	List accessible databases; filter with `like`, `startsWith`, `fromName`, `showLimit` and `history`.	A collection of Snowflake Database resources or an async result handle.
`fetchDatabase`	Fetch one database by `name`.	A Snowflake Database resource.
`createDatabase`	Create a database from a complete `body`; optional `createMode` and `kind`.	The created Database resource or async result handle.
`createOrAlterDatabase`	Create or fully alter `name` from a complete `body`.	The resulting Database resource or async result handle.
`undropDatabase`	Restore a dropped database by `name` when retention permits.	Snowflake operation status.
`deleteDatabase`	Delete `name`; `restrict` defaults to `true` in the connector.	Snowflake operation status.

Schemas

Instruction	Description	Returns
`listSchemas`	List schemas in `database`; supports name filters, `showLimit`, cursor and history.	A collection of Snowflake Schema resources or an async result handle.
`fetchSchema`	Fetch `database`.`name`.	A Snowflake Schema resource.
`createSchema`	Create a schema in `database` from a complete `body`.	The created Schema resource or async result handle.
`createOrAlterSchema`	Create or fully alter `database`.`name` from `body`.	The resulting Schema resource or async result handle.
`undropSchema`	Restore a dropped `database`.`name` when retention permits.	Snowflake operation status.
`deleteSchema`	Delete `database`.`name`; `restrict` defaults to `true` in the connector.	Snowflake operation status.

Virtual warehouses

Instruction	Description	Returns
`listWarehouses`	List warehouses, optionally filtered by `like`.	A collection of Snowflake Warehouse resources or an async result handle.
`fetchWarehouse`	Fetch one warehouse by `name`.	A Snowflake Warehouse resource.
`createWarehouse`	Create a warehouse from a complete `body`; optional `createMode`.	The created Warehouse resource or async result handle.
`createOrAlterWarehouse`	Create or fully alter `name` from `body`.	The resulting Warehouse resource or async result handle.
`resumeWarehouse`	Resume `name`; this starts compute billing.	Snowflake operation status.
`suspendWarehouse`	Suspend `name`; this affects workload availability.	Snowflake operation status.
`deleteWarehouse`	Delete `name`; optional `ifExists` makes the operation idempotent.	Snowflake operation status.

Snowflake Resource API operations can return 202 Accepted. Use the returned result handler with fetchResult until the operation completes.

DSUL Examples

Run a parameterized aggregate query:

- Snowflake.submitStatement:
    statement: |
      SELECT REGION, SUM(REVENUE) AS TOTAL
      FROM ANALYTICS.SALES
      WHERE SALE_DATE >= ?
      GROUP BY REGION
      ORDER BY TOTAL DESC
    warehouse: ANALYTICS_WH
    role: PRISME_AGENT_ROLE
    bindings:
      "1":
        type: TEXT
        value: '{{startDate}}'
  output: queryResult

Submit asynchronous SQL and poll it:

- Snowflake.submitStatement:
    statement: '{{sql}}'
    async: true
    requestId: '{{requestId}}'
  output: submitted
- Snowflake.getStatementStatus:
    statementHandle: '{{submitted.statementHandle}}'
  output: status

Create a reporting schema if it does not exist:

- Snowflake.createSchema:
    database: ANALYTICS
    createMode: ifNotExists
    body:
      name: REPORTING
      comment: Curated reporting layer
  output: schema

Inspect and suspend an idle warehouse:

- Snowflake.fetchWarehouse:
    name: ANALYTICS_WH
  output: warehouse
- conditions:
    '{{warehouse.state}} == "STARTED" and {{shouldSuspend}}':
      - Snowflake.suspendWarehouse:
          name: ANALYTICS_WH
          ifExists: true
        output: suspended

Error Handling

HTTP code	Meaning
`400`	Invalid account URL, SQL, identifier, request body or OAuth configuration.
`401`	Missing, expired, revoked or invalid Snowflake credential. OAuth users may need to reconnect.
`403`	The Snowflake role lacks the required privilege, the OAuth role is blocked, or a network policy rejects the request.
`404`	The requested statement, database, schema, warehouse or result handler is absent or not visible.
`409`	The resource already exists or its current state conflicts with the requested operation.
`422`	Snowflake rejected the resource definition or an action for its current state.
`429`	Snowflake concurrency or rate limit reached; retry with backoff.
`500` / `503`	Transient Snowflake or connector failure; retry before changing credentials.

Common Issues

“This agent is not authorized to use this connector” — Open the configuration app, add the agent under Authorized agents (or enable Allow all agents) and save. “The calling agent could not be identified” — The capability scope is missing agent_id. Set it to context_id,agent_id,user_id, then authorize the agent. “Snowflake OAuth is not connected for this user” — Open the configuration app and click Connect, or follow the connection URL surfaced by Agent Factory. “Snowflake token refresh failed … must reconnect” — The refresh token expired or was revoked. Reconnect the user. Snowflake OAuth refresh-token validity is controlled by the security integration. “Snowflake platform OAuth is not configured for an account” — Ask the platform maintainer to configure the central client, or select tenant OAuth, key-pair, programmatic-token or access-token mode. SQL output is truncated — MCP responses are intentionally limited to 100 rows and 80 KB. Narrow the query with selected columns, WHERE, aggregation and deterministic keyset pagination. A provider partition cannot recover rows omitted locally by the connector.

External Resources

Snowflake REST APIs

Official Snowflake REST API reference for databases, schemas, warehouses and asynchronous results.

Tool Agents

Learn how Agent Factory agents consume MCP tools in Prisme.ai.

Overview

App Store

SQL and results

Database catalog

Virtual warehouses

Who is this for?

Agent builder

Platform admin

Workspace builder

Prerequisites (Snowflake side)

1. Configure the connector

2. Declare the capability in AI Governance

Agent builder

Option A — Enable the shared capability from the catalog

Option B — Run it from your own workspace (recommended)

Brief the agent in its system prompt

Available Tools

Output Formats

Tool Details

sql

results

databases

schemas

warehouses

Workspace builder

Installation

Configuration

Available Instructions

SQL and asynchronous results

Databases

Schemas

Virtual warehouses

DSUL Examples

Error Handling

Common Issues

External Resources

Snowflake REST APIs

Tool Agents

SQL and results

Database catalog

Virtual warehouses

​Who is this for?

Agent builder

Platform admin

Workspace builder

​Prerequisites (Snowflake side)

​1. Configure the connector

​2. Declare the capability in AI Governance

​Agent builder

​Option A — Enable the shared capability from the catalog

​Option B — Run it from your own workspace (recommended)

​Brief the agent in its system prompt

​Available Tools

​Output Formats

​Tool Details

​sql

​results

​databases

​schemas

​warehouses

​Workspace builder

​Installation

​Configuration

​Available Instructions

​SQL and asynchronous results

​Databases

​Schemas

​Virtual warehouses

​DSUL Examples

​Error Handling

​Common Issues

​External Resources

Snowflake REST APIs

Tool Agents

Who is this for?

Prerequisites (Snowflake side)

1. Configure the connector

2. Declare the capability in AI Governance

Agent builder

Option A — Enable the shared capability from the catalog

Option B — Run it from your own workspace (recommended)

Brief the agent in its system prompt

Available Tools

Output Formats

Tool Details

sql

results

databases

schemas

warehouses

Workspace builder

Installation

Configuration

Available Instructions

SQL and asynchronous results

Databases

Schemas

Virtual warehouses

DSUL Examples

Error Handling

Common Issues

External Resources