Skyflow MCP Demo (End-to-End)

This demo shows an MCP Gateway server that:

• De-identifies PII using Skyflow Detect (string),
• Calls mocked domain sources (CRM, Hospital, Insurance, Claims),
• Drafts a response with tokens,
• Re-identifies selected fields back to plaintext.

You can run it interactively with the MCP Inspector UI and call tools like:

• crm_lookupMember, hospital_getBilling, insurance_getPolicy, claims_checkStatus
• detect_deidentify_string, skyflow_reidentify_string
• run_full_demo (one-shot final answer)
• run_full_demo_trace (full step-by-step transcript)

---

✨ What’s in here

• server_gateway.py — MCP server exposing tools and calling Skyflow APIs.
• dev_demo_skyflow.py — Single-file, non-MCP script that runs the whole flow (optional).
• (Optional) quick_client.py — Minimal Python CLI to call a tool without the UI.

---

1) Prerequisites

• Python 3.10+ (works with 3.11/3.12/3.13)
• A Skyflow Vault URL, Vault ID, Account ID, and Bearer token
• Internet access from your machine

---

2) Install

python -m venv env
source env/bin/activate
pip install -U pip
pip install httpx "mcp[cli]" python-dotenv

mcp[cli] installs the MCP Inspector (mcp dev) we’ll use as the UI.

---

3) Configure environment

Export these vars (replace with your values):

export VAULT_URL="<vault_URL>"
export VAULT_ID="<your_vault_id>"
export SKYFLOW_TOKEN="<your_bearer_token>"
export SKYFLOW_ACCOUNT_ID="<your_account_id>"

Tokens expire; if you see 401/400 errors, refresh SKYFLOW_TOKEN.

You can also put them in a .env file and source .env.

---

4) Start the MCP Server

python server_gateway.py

This starts a stdio MCP server (it won’t print output—it waits for clients).

---

5) Open the UI (MCP Inspector)

In a new terminal (keep the server running in the first one):

mcp dev server_gateway.py

This launches MCP Inspector in your browser. It will:

• Start/attach to the server for you
• Show tabs for Tools, Resources, Prompts, etc.

---

6) Use the Tools in the UI

A. Quick sanity check (string de-id)

1. Go to Tools → click detect_deidentify_string.
2. Enter text: Hi, I am Devanshu → Run Tool.
3. You should see processed_text with a token like [NAME_GIVEN_XXXX].

B. Call mocked domain tools (de-id on the fly)

Try these in Tools:

• crm_lookupMember with:

  {"member_id": "SM123456789"}
  

• insurance_getPolicy with:

  {"policy_number": "SILV-P-778845"}
  

• claims_checkStatus with:

  {"claim_id": "CLM-987654"}
  

• hospital_getBilling with:

  {"mrn": "MRN20250718001"}
  

Each returns a processed_text containing Skyflow tokens plus an entities array.

C. One-shot answer (final text only)

Use run_full_demo:

• Input:

  {"question": "Why was my SilverMoat CT claim denied?"}
  

• Output: a single user-ready sentence (re-identified where allowed).

D. Full transcript (perfect for demos)

Use run_full_demo_trace:

• Input:

  {"question": "Why was my SilverMoat CT claim denied?"}
  

• Output (JSON):
  • steps[] — each tool call with input_text, processed_text (tokens), entities, and any error
  • draft — Step 8 LLM draft (still has tokens)
  • final_text — Step 9 re-identified answer

You can expand/collapse the Structured Content pane to show everything live during a presentation.

---

7) Troubleshooting

• Server shows “no response”
  That’s normal: it’s waiting for an MCP client (Inspector or CLI). Use mcp dev server_gateway.py.

• HTTP 400 from Skyflow
  Usually input formatting or credentials. Verify:

  • VAULT_URL, VAULT_ID, SKYFLOW_ACCOUNT_ID match your tenant
  • SKYFLOW_TOKEN is fresh (tokens expire)
  • Try a simple string with detect_deidentify_string to confirm auth

• HTTP 401/403
  Token expired or wrong account/vault mapping. Refresh SKYFLOW_TOKEN.

• I want to keep emails redacted, show names
  Use the format in re-identify. The demo tools set:

  {
    "plaintext": ["name", "name_given", "name_family"],
    "redacted": ["email_address"],
    "masked": []
  }