agentsystems-notary 0.4.4

Cryptographic notarization SDK for LLM interactions

AgentSystems Notary

Audit logging infrastructure for AI systems

AgentSystems Notary provides tamper-evident audit trails for AI systems. It creates cryptographically verifiable logs of all LLM interactions with dual-write architecture: your storage bucket (raw logs) + hash storage (verification receipts).

Features

• Multi-Framework Support: LangChain and CrewAI adapters
• Dual-Write Architecture: Your bucket (raw logs) + hash storage (receipts)
• Flexible Hash Storage: Arweave (decentralized) and/or Custodied (AgentSystems API)
• Cryptographic Verification: SHA-256 hashes with JCS canonicalization (RFC 8785)
• Multi-Tenant Support: Isolated audit trails for SaaS applications

Installation

pip install agentsystems-notary

Quick Start

Copy .env.example to .env and fill in your credentials.

LangChain

pip install langchain-anthropic

import os
from dotenv import load_dotenv
from agentsystems_notary import (
    LangChainNotary,
    RawPayloadStorage,
    ArweaveHashStorage,
    LocalKeySignerConfig,
    AwsS3StorageConfig,
)
from langchain_anthropic import ChatAnthropic

load_dotenv()

# Where full audit payloads are stored (your S3 bucket)
raw_payload_storage = RawPayloadStorage(
    storage=AwsS3StorageConfig(
        bucket_name=os.environ["ORG_AWS_S3_BUCKET_NAME"],
        aws_access_key_id=os.environ["ORG_AWS_S3_ACCESS_KEY_ID"],
        aws_secret_access_key=os.environ["ORG_AWS_S3_SECRET_ACCESS_KEY"],
        aws_region=os.environ["ORG_AWS_S3_REGION"],
    ),
)

# Where hashes are stored — Arweave for independent verification
hash_storage = [
    ArweaveHashStorage(
        namespace="my_namespace",
        signer=LocalKeySignerConfig(
            private_key_path=os.environ["ARWEAVE_PRIVATE_KEY_PATH"],
        ),
        bundler_url=os.environ["ARWEAVE_BUNDLER_URL"],
    ),
]

# Initialize notary
notary = LangChainNotary(
    raw_payload_storage=raw_payload_storage,
    hash_storage=hash_storage,
    debug=True,
)

model = ChatAnthropic(
    model="claude-sonnet-4-5-20250929",
    api_key=os.environ["ANTHROPIC_API_KEY"],
    callbacks=[notary],
)

response = model.invoke("What is 2 + 2?")

CrewAI

pip install crewai

import os
from dotenv import load_dotenv
from agentsystems_notary import (
    CrewAINotary,
    RawPayloadStorage,
    ArweaveHashStorage,
    LocalKeySignerConfig,
    AwsS3StorageConfig,
)
from crewai import Agent, Task, Crew, LLM

load_dotenv()

# Where full audit payloads are stored (your S3 bucket)
raw_payload_storage = RawPayloadStorage(
    storage=AwsS3StorageConfig(
        bucket_name=os.environ["ORG_AWS_S3_BUCKET_NAME"],
        aws_access_key_id=os.environ["ORG_AWS_S3_ACCESS_KEY_ID"],
        aws_secret_access_key=os.environ["ORG_AWS_S3_SECRET_ACCESS_KEY"],
        aws_region=os.environ["ORG_AWS_S3_REGION"],
    ),
)

# Where hashes are stored — Arweave for independent verification
hash_storage = [
    ArweaveHashStorage(
        namespace="my_namespace",
        signer=LocalKeySignerConfig(
            private_key_path=os.environ["ARWEAVE_PRIVATE_KEY_PATH"],
        ),
        bundler_url=os.environ["ARWEAVE_BUNDLER_URL"],
    ),
]

# Initialize notary (hooks register automatically)
notary = CrewAINotary(
    raw_payload_storage=raw_payload_storage,
    hash_storage=hash_storage,
    debug=True,
)

llm = LLM(
    model="anthropic/claude-sonnet-4-5-20250929",
    api_key=os.environ["ANTHROPIC_API_KEY"],
)
agent = Agent(role="Analyst", goal="Answer questions", backstory="Expert analyst", llm=llm)
task = Task(description="What is 2 + 2?", expected_output="The answer", agent=agent)
crew = Crew(agents=[agent], tasks=[task])

result = crew.kickoff()

How It Works

1. Capture: Intercepts LLM requests/responses via framework hooks
2. Canonicalize: Deterministic JSON serialization (JCS/RFC 8785)
3. Hash: SHA-256 of canonical bytes
4. Dual-Write:
   • Your bucket: Full canonical JSON payload
   • Hash storage: Hash receipt for verification

Configuration

Raw Payload Storage

Where full audit payloads are stored (your bucket):

from agentsystems_notary import RawPayloadStorage, AwsS3StorageConfig

raw_payload_storage = RawPayloadStorage(
    storage=AwsS3StorageConfig(
        bucket_name="my-audit-logs",
        aws_access_key_id="...",
        aws_secret_access_key="...",
        aws_region="us-east-1",
    ),
)

Hash Storage

Where hashes are stored for verification. You can use one or both.

Arweave (Decentralized) — Public blockchain, permanent storage, no vendor dependency. Verify independently with open-source CLI.

from agentsystems_notary import ArweaveHashStorage, LocalKeySignerConfig

ArweaveHashStorage(
    namespace="my_namespace",
    signer=LocalKeySignerConfig(
        private_key_path="path/to/rsa-4096-private.pem",
    ),
    bundler_url="https://node2.bundlr.network",
)

Custodied (AgentSystems API) — Managed service if you prefer AgentSystems to handle the complexity.

from agentsystems_notary import CustodiedHashStorage

CustodiedHashStorage(
    api_key="sk_asn_prod_...",  # From agentsystems.ai
    slug="my_tenant",
)

Using both:

hash_storage=[
    ArweaveHashStorage(namespace="my_namespace", signer=..., bundler_url="..."),
    CustodiedHashStorage(api_key="...", slug="my_tenant"),
]

Debug Mode

notary = LangChainNotary(
    raw_payload_storage=...,
    hash_storage=[...],
    debug=True,  # Prints canonical JSON and hashes
)

S3 Bucket Structure

{env}/{namespace}/{YYYY}/{MM}/{DD}/{hash}.json

• env: arweave, prod, or test
• namespace: Your namespace (Arweave) or tenant ID from API response (custodied)
• hash: SHA-256 hash of the canonical payload

Verification

For Arweave-notarized logs, use the open-source CLI — no account required:

npm install -g agentsystems-verify
agentsystems-verify --logs logs.zip

Manual verification:

import hashlib

# 1. Download payload from your bucket
with open("payload.json", "rb") as f:
    canonical_bytes = f.read()

# 2. Compute hash
computed_hash = hashlib.sha256(canonical_bytes).hexdigest()

# 3. Compare with stored hash (from Arweave or custodied receipt)
assert computed_hash == stored_hash

Support

• Documentation: docs.agentsystems.ai/notary
• Dashboard: notary.agentsystems.ai
• Issues: GitHub Issues

License

Licensed under the Apache-2.0 license.