LiteLLM Integration

The Glacis LiteLLM integration wraps LiteLLM’s unified completion interface to automatically create cryptographic attestations for every LLM call across 100+ providers. Your data is hashed locally and never leaves your environment — only hashes and metadata are sent to the Glacis transparency log.

Since LiteLLM exposes module-level functions (not a client object), this integration returns a thin wrapper object with .completion() and .acompletion() methods that mirror LiteLLM’s interface.

Installation

pip install glacis[litellm]

Quick Start

from glacis.integrations.litellm import attested_litellm, get_last_receipt

client = attested_litellm(glacis_api_key="glsk_live_...")

# Make a normal LiteLLM call — attestation happens automatically
response = client.completion(
    model="gpt-4",
    messages=[{"role": "user", "content": "Hello!"}]
)

print(response.choices[0].message.content)

# Get the attestation receipt
receipt = get_last_receipt()
print(f"Attested: {receipt.id}")
print(f"Status: {receipt.witness_status}")

Note

The wrapper accepts the same arguments as litellm.completion(). Replace litellm.completion(...) with client.completion(...) and attestation happens automatically.

Note

All calls through this integration are recorded with operation_type="completion". For custom operation types, use glacis.attest() directly.

What Gets Attested

For each completion, Glacis captures:

Field	Treatment	Details
Request messages	Hashed	SHA-256, never sent to Glacis
Response content	Hashed	SHA-256, never sent to Glacis
System prompt	Hashed	SHA-256 hash included in control plane record
Model name	Metadata	Sent as-is (e.g., gpt-4, claude-3-sonnet, mistral/mistral-large)
Temperature	Metadata	Included in control plane record
Token counts	Metadata	prompt, completion, and total tokens
Provider	Metadata	Always "litellm"

Usage with Different Providers

LiteLLM supports 100+ providers through a unified interface. The model parameter determines which provider is used:

from glacis.integrations.litellm import attested_litellm

client = attested_litellm(glacis_api_key="glsk_live_...")

# OpenAI
response = client.completion(model="gpt-4", messages=[...])

# Anthropic
response = client.completion(model="claude-3-sonnet-20240229", messages=[...])

# Azure OpenAI
response = client.completion(model="azure/my-deployment", messages=[...])

# AWS Bedrock
response = client.completion(model="bedrock/anthropic.claude-v2", messages=[...])

# Ollama (local)
response = client.completion(model="ollama/llama3", messages=[...])

Each call is attested with the actual model name used, giving you a unified attestation trail across all your LLM providers.

Environment Variables

Online

from glacis.integrations.litellm import attested_litellm

# LiteLLM reads provider API keys from env vars automatically
# (OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.)
client = attested_litellm(glacis_api_key="glsk_live_...")

Offline

import os
from glacis.integrations.litellm import attested_litellm

# No Glacis API key needed for offline mode
client = attested_litellm(
    offline=True,
    signing_seed=os.urandom(32),
)

Accessing Receipts

Use get_last_receipt() to retrieve the attestation from the most recent API call. Receipts are stored in thread-local storage, so each thread maintains its own last receipt independently.

from glacis.integrations.litellm import get_last_receipt

receipt = get_last_receipt()
if receipt:
    print(f"ID: {receipt.id}")
    print(f"Evidence hash: {receipt.evidence_hash}")
    print(f"Status: {receipt.witness_status}")  # "WITNESSED" or "UNVERIFIED"
    print(f"Service: {receipt.service_id}")

Caution

get_last_receipt() returns the receipt for the current thread only. In multi-threaded applications, each thread tracks its own last receipt. Call get_last_receipt() immediately after the API call to avoid it being overwritten by a subsequent call.

Offline Mode

Offline mode creates locally-signed attestations without connecting to the Glacis server. This is useful for development, air-gapped environments, or when you want to defer attestation submission.

import os
from glacis.integrations.litellm import attested_litellm, get_last_receipt

client = attested_litellm(
    offline=True,
    signing_seed=os.urandom(32),
)

response = client.completion(
    model="gpt-4",
    messages=[{"role": "user", "content": "Hello!"}]
)

receipt = get_last_receipt()
print(f"Status: {receipt.witness_status}")  # "UNVERIFIED"

Tip

See the Offline Mode page for details on offline signing, seed management, and later submission.

Using Controls

Controls let you scan inputs and outputs for PII, jailbreak attempts, banned words, and more. Configure them via a glacis.yaml file:

from glacis.integrations.litellm import attested_litellm, GlacisBlockedError

client = attested_litellm(config="glacis.yaml")

try:
    response = client.completion(
        model="gpt-4",
        messages=[{"role": "user", "content": "Hello!"}]
    )
    print(response.choices[0].message.content)
except GlacisBlockedError as e:
    print(f"Blocked by {e.control_type} (score={e.score})")

Caution

Input controls scan only the last user message in the conversation. Earlier messages in multi-turn conversations are not scanned. If you need to scan all messages, run controls manually before calling the integration.

Custom Metadata

Attach business-specific metadata to every attestation created by the client:

from glacis.integrations.litellm import attested_litellm, get_last_receipt

client = attested_litellm(
    glacis_api_key="glsk_live_...",
    metadata={"department": "legal", "use_case": "contract-review"},
)

response = client.completion(
    model="gpt-4",
    messages=[{"role": "user", "content": "Summarize this contract."}],
)

# Metadata is included in every attestation:
# {"provider": "litellm", "model": "gpt-4", "department": "legal", "use_case": "contract-review"}

Note

Metadata values must be strings. The reserved keys provider and model are set automatically and cannot be overridden.

Retrieving Evidence

Evidence includes the full input, output, and control plane results that were attested. Evidence is stored locally and never sent to Glacis servers.

from glacis.integrations.litellm import get_last_receipt, get_evidence

receipt = get_last_receipt()
if receipt:
    evidence = get_evidence(receipt.id)
    if evidence:
        print(evidence["input"])    # Original request (model, messages)
        print(evidence["output"])   # Full response (choices, usage)

Async Support

The wrapper also provides acompletion() for async workflows:

import asyncio
from glacis.integrations.litellm import attested_litellm, get_last_receipt

async def main():
    client = attested_litellm(glacis_api_key="glsk_live_...")

    response = await client.acompletion(
        model="gpt-4",
        messages=[{"role": "user", "content": "Hello!"}]
    )

    print(response.choices[0].message.content)
    receipt = get_last_receipt()
    print(f"Attested: {receipt.id}")

asyncio.run(main())

attested_litellm() Reference

Parameter	Type	Default	Description
glacis_api_key	Optional[str]	None	Glacis API key. Required for online mode.
glacis_base_url	str	"https://api.glacis.io"	Glacis API base URL.
service_id	str	"litellm"	Service identifier for attestations.
debug	bool	False	Enable debug logging.
offline	Optional[bool]	None	Enable offline mode. If None, inferred from config or presence of glacis_api_key.
signing_seed	Optional[bytes]	None	32-byte Ed25519 signing seed. Required when offline=True.
policy_key	Optional[bytes]	None	32-byte HMAC key for sampling decisions. Falls back to signing_seed if not provided.
config	Optional[str]	None	Path to glacis.yaml config file for controls, sampling, and policy settings.
input_controls	Optional[list[BaseControl]]	None	Custom controls to run on input text before the LLM call.
output_controls	Optional[list[BaseControl]]	None	Custom controls to run on output text after the LLM call.
metadata	Optional[dict[str, str]]	None	Custom metadata to include in every attestation. Merged with provider defaults (provider, model). Cannot override reserved keys.

Returns: An AttestedLiteLLM wrapper with .completion() and .acompletion() methods.

Raises: GlacisBlockedError if a control blocks the request.

Streaming

Note

Streaming is not yet supported. Passing stream=True may produce errors or incomplete attestations, as the wrapper expects a non-streaming response object. Streaming support is planned for a future release.