langchain-nobulex

Compliance middleware for LangChain agents with hash-chained audit trails.

Enforce covenant rules on agent tool calls, block forbidden actions, and maintain a tamper-proof audit log — ready for EU AI Act compliance requirements.

Installation

pip install langchain-nobulex

Quick Start

from langchain_nobulex import create_agent, NobulexComplianceMiddleware

agent = create_agent(
    model="gpt-4.1",
    tools=tools,
    middleware=[
        NobulexComplianceMiddleware(
            rules="permit read; forbid transfer where amount > 500; require log_all;"
        )
    ]
)

How It Works

Define rules using a simple covenant DSL:

• permit <action> — explicitly allow a tool/action
• forbid <action> [where <condition>] — block a tool/action, optionally with conditions
• require <behavior> — declare required behaviors (e.g., log_all)

The middleware:

1. Intercepts every tool call from the model
2. Matches it against your covenant rules
3. Blocks forbidden actions before they execute
4. Logs every action (allowed and blocked) in a SHA-256 hash-chained audit trail

Verify integrity independently:

from langchain_nobulex import verify

log = agent.get_audit_log()
assert verify(log)  # True if chain is intact, False if tampered

Audit Trail

Every entry in the audit log contains:

Each hash links to the previous entry, forming a tamper-evident chain. Use verify() to independently validate the entire chain.

EU AI Act Compliance

This middleware supports EU AI Act (Regulation 2024/1689) requirements for high-risk AI systems:

• Article 14 — Human oversight: Rules let operators define boundaries before deployment
• Article 12 — Record-keeping: Hash-chained audit logs provide tamper-proof records of all AI actions
• Article 9 — Risk management: Forbidden rules act as runtime guardrails against identified risks
• Independent verification: The verify() function enables third-party audit of the action log

API Reference

NobulexComplianceMiddleware(rules: str)

Main middleware class. Pass a covenant rules string.

• after_model(response) — intercepts and filters tool calls
• wrap_tool_call(tool_call) — hook for per-execution logging
• after_agent(state) — attaches audit log to final agent state
• get_audit_log() — returns the full audit trail as a list of dicts
• get_audit_log_json() — returns the audit trail as formatted JSON

verify(audit_log)

Independently validates a hash-chained audit log. Returns True if intact.

create_agent(model, tools, middleware)

Creates a middleware-aware agent wrapper.

License

MIT