Skip to content

The future of AI, revealed live

Stream Summit keynotes free June 1–2.

Register now
Snowflake for Developers/Guides/Integrating Cortex API with Cursor Agent
Quickstart

Integrating Cortex API with Cursor Agent

Cortex LLM
Priya Joseph
Fork Repo

A quickstart for routing the Cursor Agent's chat completions to Snowflake Cortex (claude-4-sonnet) via a local LiteLLM proxy.

A local proxy pattern that exposes Snowflake Cortex's REST API as an OpenAI-compatible endpoint so any OpenAI-compatible coding agent can use it as a backend.

Cursor configured with a custom Cortex model

Why a proxy?

Cursor's "Add Models" dialog only accepts OpenAI-compatible endpoints (POST /v1/chat/completions). Snowflake Cortex's REST API uses its own request/response schema (POST /api/v2/cortex/inference:complete with header X-Snowflake-Authorization-Token-Type). LiteLLM has a built-in snowflake/ provider that translates between the two, so we run it locally as a 1-command proxy.

Cursor Agent ── OpenAI /v1/chat/completions ──▶ LiteLLM (localhost:4000) ──▶ Snowflake Cortex /api/v2/cortex/inference:complete

Prerequisites

  • Cursor (Pro / Business / Enterprise — see "Team API key restriction" at the bottom).
  • Python 3.10+ available locally.
  • A Snowflake account with Cortex enabled in your region. This guide was verified on AWS_US_WEST_2 with model claude-4-sonnet.
  • A Snowflake Programmatic Access Token (PAT) scoped to a role that has USAGE on Cortex. Generate from Snowsight → your profile → Programmatic access tokens.

1. Verify Cortex availability

In a Snowflake worksheet:

USE WAREHOUSE <your_wh>;
SELECT CURRENT_REGION();
SELECT SNOWFLAKE.CORTEX.COMPLETE('claude-4-sonnet', 'ping');

If claude-4-sonnet is unavailable in your region, swap to another supported Claude model (e.g. claude-3-5-sonnet, claude-3-7-sonnet) and use that name everywhere below.

2. Smoke-test the REST endpoint

ACCOUNT_HOST="<youraccount>.snowflakecomputing.com"   
PAT="<your_pat>"

curl -sS https://$ACCOUNT_HOST/api/v2/cortex/inference:complete \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $PAT" \
  -H "X-Snowflake-Authorization-Token-Type: PROGRAMMATIC_ACCESS_TOKEN" \
  -d '{"model":"claude-4-sonnet","messages":[{"role":"user","content":"ping"}],"max_tokens":20}'

Expect a data:-prefixed SSE stream ending with a usage block. If you get 401 / 403, double-check the PAT and role grants.

3. Install LiteLLM in an isolated venv

mkdir -p ~/cursor-cortex && cd ~/cursor-cortex
python3 -m venv .venv
.venv/bin/pip install --upgrade pip
.venv/bin/pip install 'litellm[proxy]'
.venv/bin/litellm --version   # tested on 1.83.14

4. Author litellm_config.yaml

model_list:
  - model_name: cortex-claude-4-sonnet
    litellm_params:
      model: snowflake/claude-4-sonnet
      api_base: https://<youraccount>.snowflakecomputing.com/api/v2
      api_key: os.environ/SNOWFLAKE_PAT      # MUST be prefixed with `pat/` — see below
      stream: true

litellm_settings:
  drop_params: true
  num_retries: 2
  request_timeout: 120

general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY
  disable_spend_logs: true

Important: LiteLLM's Snowflake provider defaults to KEYPAIR_JWT auth. To use a PAT, the API key must literally start with pat/. We do this in the .env file below.

5. Create .env (mode 0600)

cat > .env <<EOF
SNOWFLAKE_PAT=pat/<your_pat>
LITELLM_MASTER_KEY=sk-cursor-$(python3 -c 'import secrets;print(secrets.token_urlsafe(24))')
EOF
chmod 600 .env

6. Create run-proxy.sh

#!/bin/bash
set -euo pipefail
DIR="$(cd "$(dirname "$0")" && pwd)"
set -a; source "$DIR/.env"; set +a
exec "$DIR/.venv/bin/litellm" \
  --config "$DIR/litellm_config.yaml" \
  --port 4000 --host 127.0.0.1

chmod +x run-proxy.sh
./run-proxy.sh    # leave running; logs stream to stdout

Verify:

source .env
curl -sS http://127.0.0.1:4000/v1/models -H "Authorization: Bearer $LITELLM_MASTER_KEY"
curl -sS http://127.0.0.1:4000/v1/chat/completions \
  -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"cortex-claude-4-sonnet","messages":[{"role":"user","content":"ping"}]}'

You should see standard OpenAI-shape JSON.

7. Configure Cursor (3.x)

  1. In any Cursor chat, click the model picker (e.g. composer-2-fast).
  2. At the bottom of the popup, click Add Models.
  3. Choose OpenAI (or "OpenAI-compatible") and enter:
    • Base URL: http://127.0.0.1:4000/v1
    • API Key: the value of LITELLM_MASTER_KEY from .env
    • Model name: cortex-claude-4-sonnet
  4. Save / Verify. Select cortex-claude-4-sonnet in the picker.

8. Verify end-to-end

  • Send a chat message in Cursor → request appears in the LiteLLM stdout log.

  • After ~1–2h, confirm token usage in Snowflake:

    SELECT MODEL_NAME, SUM(TOKENS) AS tokens
FROM SNOWFLAKE.ACCOUNT_USAGE.CORTEX_REST_API_USAGE_HISTORY
WHERE START_TIME >= DATEADD(hour, -3, CURRENT_TIMESTAMP())
GROUP BY 1;

Team API key restriction

If Cursor shows:

Your team has disabled OpenAI API keys. Please disable your API key in Cursor Settings > Models.

…your Cursor admin has turned off BYOK at the team level. Options:

  • Ask the admin to enable "Allow custom OpenAI API keys" in the Cursor team console.
  • Or use the proxy from a personal/non-team Cursor account that doesn't inherit the policy.
  • Or, instead of Cursor, point any OpenAI-compatible client (Continue.dev, Aider, Open WebUI, custom scripts) at the same http://127.0.0.1:<port>/v1 proxy — they have no such restriction.

Troubleshooting

SymptomCauseFix
401 Unauthorized from CortexPAT missing or expiredRegenerate PAT, update .env, restart proxy
Missing Snowflake JWT key from LiteLLMSNOWFLAKE_PAT not in envset -a; source .env; set +a before launching
Proxy returns KEYPAIR_JWT auth errorForgot pat/ prefixSNOWFLAKE_PAT=pat/<token> in .env
Model "claude-x-y" is unavailableModel not in your regionSHOW FUNCTIONS LIKE 'COMPLETE%' and try a supported Claude variant
Cursor "Verify" button failsProxy not running on 127.0.0.1:<port>lsof -i:<port>; restart run-proxy.sh

File layout

~/cursor-cortex/
├── .env                  # SNOWFLAKE_PAT + LITELLM_MASTER_KEY (mode 0600)
├── .venv/                # isolated python env w/ litellm[proxy]
├── litellm_config.yaml   # model_list + master key reference
└── run-proxy.sh          # source .env && exec litellm
Updated 2026-05-13

This content is provided as is, and is not maintained on an ongoing basis. It may be out of date with current Snowflake instances