langchain-signatrust

LangChain tools for Signatrust — cryptographically signed, tamper-evident AI Decision Receipts.

Give any LangChain agent the ability to seal its high-stakes decisions into independently verifiable evidence — without exposing prompts, model outputs, or sensitive business data. Only SHA-256 hashes are stored by default.

Why?

When an AI agent makes a regulated or high-stakes decision (loan approval, refund, content moderation, transaction flagging), how do you prove, after the fact, that the decision was made correctly, under the right policies, and with appropriate human oversight?

Signatrust generates a tamper-evident AI Decision Receipt for each decision, capturing:

• Which AI system (model + version) was involved
• The action taken and decision context
• Whether a human reviewed it
• The policies and permissions in effect
• A cryptographic (Ed25519) signature proving the record has not been altered

Each receipt has a public verify_url that regulators, auditors, or counterparties can check — without accessing your systems or data.

Installation

pip install langchain-signatrust

Authentication

Get an API key (starts with sk_live_…) from signatrust.net/register, then either pass it explicitly or set an environment variable:

export SIGNATRUST_API_KEY="sk_live_..."
# Optional, for self-hosted Enterprise:
export SIGNATRUST_BASE_URL="https://signatrust.your-company.com/api/v1"

Quick start

from langchain_signatrust import SignatrustGenerateReceiptTool

tool = SignatrustGenerateReceiptTool()  # reads SIGNATRUST_API_KEY from env

receipt = tool.invoke({
    "agent_name": "LoanApprovalAgent",
    "action": "Approved loan application #4821",
    "decision": "APPROVED: applicant meets internal credit policy v3",
    "risk_level": "high",
    "human_review": True,
    "model_provider": "openai",
    "model_name": "gpt-4o",
    "policies": ["eu-ai-act-high-risk", "internal-credit-v3"],
    "permissions": ["credit.decide"],
})

print(receipt["receipt_id"])   # e.g. STR-1A2B3C4D5E
print(receipt["verify_url"])   # https://verify.signatrust.net/r/STR-1A2B3C4D5E

Use with an agent

from langchain_signatrust import get_signatrust_tools
from langchain_openai import ChatOpenAI
from langgraph.prebuilt import create_react_agent

tools = get_signatrust_tools()           # all three tools at once
llm = ChatOpenAI(model="gpt-4o")
agent = create_react_agent(llm, tools)

agent.invoke({"messages": [
    ("user", "Approve the refund for order #991 and generate a signed decision receipt.")
]})

Tools provided

All three tools accept api_key, base_url, and timeout constructor arguments; if omitted, they fall back to the SIGNATRUST_API_KEY / SIGNATRUST_BASE_URL environment variables.

Privacy-first by design

By default, only the SHA-256 hash of the decision and input_prompt is stored — never the raw text. Pass include_decision_in_metadata=True only if you explicitly want to retain the raw decision in the receipt metadata.

Development

pip install -e ".[test]"
pytest

Links

• Website: signatrust.net
• API docs: signatrust.net/docs/api
• Source: github.com/abokenan444/Signatrust
• Contact: partners@signatrust.net

© 2026 Signatrust — MIT License