❄️ snowflake-claude-code

Run Claude Code against Claude models served by Cortex inside your Snowflake account, so prompts and responses never leave your Snowflake governance boundary.

Claude Code  →  FastAPI proxy (127.0.0.1:4000)  →  Snowflake Cortex Inference

The CLI authenticates to Snowflake (SSO or PAT), starts a local proxy that translates Anthropic Messages API calls to Cortex Inference calls (SSE streaming included), and launches Claude Code pointed at the proxy.

⚡ Quick start

uv tool install snowflake-claude-code
npm install -g @anthropic-ai/claude-code
snowflake-claude-code --account MYORG-MYACCOUNT --user me@company.com

Browser pops for Snowflake SSO, proxy spins up, Claude Code launches.

🔒 Why

Your Claude Code session never talks to Anthropic. Every prompt, file read, tool call, and model response goes over TLS to the same Snowflake endpoint your warehouse queries already use — governed by your existing Snowflake trust boundary, not a new third-party LLM vendor.

• 🚫 No traffic to Anthropic. The proxy binds to 127.0.0.1 only; the only outbound endpoint is your Snowflake account's API.
• 🛡️ Snowflake IAM applies. Role, warehouse, and network policy controls gate model access. Revoke Snowflake → revoke AI.
• 🔑 Familiar auth. Browser SSO flows through your existing IdP; PATs for headless.
• 📝 Full audit trail. Every call lands in SNOWFLAKE.ACCOUNT_USAGE.CORTEX_REST_API_USAGE_HISTORY.
• 🌍 Data residency honored. Inference runs in your account's region.
• 🧠 No training on your data. Per Snowflake Cortex terms.
• 💰 Consolidated spend. Cortex tokens roll up with your warehouse costs.
• ♻️ Transparent re-auth. Expired tokens trigger a silent refresh mid-session.

📦 Install

Requires Python 3.10+ and the Claude Code CLI.

uv tool install snowflake-claude-code     # recommended
pipx install snowflake-claude-code        # or pipx
pip install snowflake-claude-code         # or pip
uvx snowflake-claude-code ...             # or run without installing

And Claude Code itself:

npm install -g @anthropic-ai/claude-code

🚀 Usage

snowflake-claude-code \
  --account MYORG-MYACCOUNT \
  --user me@company.com \
  --model opus

Flags

Flag	Default	Description
--account	—	Snowflake account identifier
--user	—	Snowflake username (required)
--model	sonnet	Cortex model ID, or a family alias (opus, sonnet, haiku)
--port	4000	Local proxy port
--token	—	Snowflake PAT — pair with --user to skip browser SSO
--list-models	—	Print the Cortex models this account can reach, then exit
--verbose, -v	off	Debug logging

Environment variables

Any flag can also be provided via an env var — useful for shell profiles, CI, or devcontainers:

Env var	Equivalent flag
SNOWFLAKE_ACCOUNT	--account
SNOWFLAKE_USER (or SNOWFLAKE_USERNAME)	--user
SNOWFLAKE_MODEL	--model
SNOWFLAKE_PORT	--port
SNOWFLAKE_TOKEN	--token

export SNOWFLAKE_ACCOUNT=MYORG-MYACCOUNT
export SNOWFLAKE_USER=me@company.com
snowflake-claude-code

Config file

Or persist them in ~/.snowflake-claude-code/config.toml:

account = "MYORG-MYACCOUNT"
user = "me@company.com"
default_model = "sonnet"
port = 4000
# token = "pat-..."   # optional, skips SSO

Precedence: CLI flags > env vars > config file > defaults.

🤖 Supported models

Pass --model either a family alias or an explicit Cortex model ID.

Value	Resolves to
sonnet	Newest generally available Sonnet on your account (the default)
opus	Newest generally available Opus
haiku	Newest generally available Haiku
claude-opus-5 (or any ID)	Used verbatim — including public-preview and non-Claude models

Explicit IDs are passed through untouched, so a --model flag or default_model pinned before aliases existed keeps selecting exactly the model it names.

Aliases resolve at startup from SHOW CORTEX BASE MODELS, so they track new Cortex releases without an upgrade, and only ever pick a GA model — preview models must be named explicitly. The query is filtered to models your role holds grants on, needs no running warehouse, and costs no credits. If it fails, a built-in last-known-good list is used instead.

/v1/models advertises whatever your account can actually reach, which is what Claude Code's model picker shows. To see the same list from the terminal — and what each alias resolves to on your account — run snowflake-claude-code --list-models.

Region availability still applies — a model listed for your account may need cross-region inference to run. Non-Claude Cortex models work for plain chat (--model mistral-large2, --model llama3.1-70b); tool-calling compatibility varies.

🔍 Verify traffic is hitting Snowflake

SELECT START_TIME, MODEL_NAME, TOKENS, USER_ID, INFERENCE_REGION
FROM SNOWFLAKE.ACCOUNT_USAGE.CORTEX_REST_API_USAGE_HISTORY
WHERE START_TIME >= CURRENT_DATE()
ORDER BY START_TIME DESC;

ACCOUNT_USAGE views lag 45 min–3 hours. For real-time, run with --verbose.

🏗️ Architecture

snowflake_claude_code/
├── cli.py        Parse config, start proxy, launch `claude` subprocess
├── proxy.py      FastAPI app: /v1/messages, /v1/models, /v1/health
├── translate.py  Anthropic ⇄ Cortex format translation + SSE adapter
├── models.py     Cortex model discovery + family alias resolution
├── auth.py       Snowflake connector + re-auth on 401
└── config.py     Layered config loader

The proxy binds to 127.0.0.1 only. The Snowflake token lives in process memory for the session lifetime and is cleared on exit.

🛠️ Development

git clone https://github.com/dylan-murray/snowflake-claude-code.git
cd snowflake-claude-code
uv sync --group dev

uv run pytest
uv run ruff check .
uv run ruff format .

CI runs on Python 3.10–3.14, against the locked dependencies, the lowest declared bounds, and the newest versions on PyPI.

📄 License

MIT — see LICENSE.