litellm-shield Integration

litellm-shield (PyPI: 0din-litellm-shield) is an open-source LiteLLM guardrail plugin that runs SusFactor against every prompt — and optionally every response — passing through a LiteLLM proxy. It's a drop-in way to consume the SusFactor early-access beta from LiteLLM without writing your own client. (In-process sdk/sdk-onnx backends also exist in the package, but currently depend on a private 0DIN wheel and aren't installable outside 0DIN — external users should use backend: "hosted".)

Install

pip install "litellm[proxy]" "0din-litellm-shield"

Configure

Add a guardrails block to your LiteLLM proxy config, pointing at litellm_shield.SusFactorGuardrail with backend: "hosted":

guardrails:
  - guardrail_name: "susfactor"
    litellm_params:
      guardrail: litellm_shield.SusFactorGuardrail
      mode: "during_call"        # runs in parallel with the LLM call; hides latency
      enforcement: "flag"        # shadow | flag | block
      threshold: 0.5
      fail_open: true            # allow requests through if SusFactor is unavailable
      backend: "hosted"
      access_token: os.environ/ODIN_ACCESS_TOKEN
      default_on: true           # apply to every request automatically

Start the proxy: litellm --config config.yaml.

Credentials

access_token is the same 0din.ai Portal API key used elsewhere in these docs — litellm-shield exchanges it for a short-lived JWT and refreshes it in the background, so you only manage one long-lived credential. Set it via the ODIN_ACCESS_TOKEN environment variable (ODIN_API_TOKEN also works as a fallback).

Enforcement modes

enforcement	Suspicious prompt	Client sees
shadow	Logged only, never blocked	Nothing
flag (default)	Allowed through, annotated	X-SusFactor-Decision response header
block	Rejected	Guardrail-violation error (HTTP 400)

mode controls when scoring happens relative to the LLM call: pre_call (before dispatch — use with block if a suspicious prompt must never reach the model), during_call (default; scores in parallel with the model call), or post_call (after the response; used with scan_output: true to also score model output).

Limitations

• The hosted backend is the only one available outside 0DIN — the sdk/sdk-onnx in-process backends depend on a private wheel and aren't installable from PyPI.
• Prompts and responses scored via the hosted backend are sent to 0DIN's hosted SusFactor API, subject to the same early-access beta terms as direct API usage — endpoints and rate limits may change, and production or sensitive data shouldn't be sent.
• fail_open defaults to true — set it to false if a SusFactor outage should fail closed rather than let requests through unscored.
• The package is at an early (alpha) maturity stage per its PyPI classifiers.

Learn more

See the litellm-shield README for the full configuration reference, Docker Compose quickstart, and self-hosted backend setup.