Skip to content
 
 

Latest commit

 

History

4 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Snowflake Managed MCP HOL

Public workshop materials for a governed neobank support agent built with a Snowflake Managed MCP Server.

The lab exposes four narrowly scoped Snowflake tools:

  • classify_intent: classifies a customer message with a fine-tuned model
  • search_help_articles: searches a synthetic help-centre knowledge base
  • get_transaction_status: looks up one synthetic case by reference ID
  • file_ticket: creates and routes a support incident

All included customer records and support content are synthetic. This repository contains no credentials, account hostnames, or executed notebook output.

Prerequisites

  • A Snowflake account with Managed MCP Server, Cortex Search, and Cortex fine-tuning access
  • A warehouse named COCO_WH, or equivalent edits to the SQL and notebook
  • A role allowed to create the objects used by the lab
  • The public Banking77 dataset, included as generated SQL in sql/01_banking77.sql

Banking77 is published under CC BY 4.0. The generated tables contain TEXT and LABEL columns.

Setup

Run the scripts in this order:

  1. sql/00_setup.sql
  2. sql/00_cases_and_status.sql
  3. sql/01_banking77.sql
  4. sql/03_search_help_articles.sql
  5. sql/06_finetune_intent.sql
  6. sql/02_case_ticket_and_routing.sql
  7. sql/04_mcp_server.sql

The fine-tuning job in step 5 is asynchronous. The procedure can be created immediately, but wait for the job to succeed before calling CLASSIFY_INTENT_PROC or creating the MCP server.

To regenerate the Banking77 SQL from the publisher's CSV files, run scripts/build_banking77_sql.py TRAIN_CSV TEST_CSV sql/01_banking77.sql.

Open In Snowflake Workspaces

  1. In Snowsight, open Projects > Workspaces.
  2. Select From Git repository.
  3. Paste https://github.com/sfc-gh-zhada/mcp-hol-demo.
  4. Select Public repository. No GitHub credentials are required.
  5. Choose an API integration approved by your Snowflake administrator.
  6. Open mcp_hol_demo.ipynb.

Security Model

The MCP server exposes only the four declared tools. The CUSTOMER_AGENT role receives tool usage privileges but no direct SELECT privilege on the underlying case or ticket tables. The write action creates a support incident only; it does not move money or call an external service.

The sample creates or replaces objects in the MCP_HOL database. Use a dedicated workshop account or schema, and review object names before running it in a shared environment.

License

The workshop code in this repository is provided under the MIT License. Banking77 remains subject to its own CC BY 4.0 license and attribution requirements.

About

No description, website, or topics provided.

Resources

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages