evrguard 0.1.0

Deterministic safety for AI agents. Mathematics, not prompts.

EvrGuard

Deterministic safety for AI agents. Mathematics, not prompts.

---

What is EvrGuard?

EvrGuard wraps any AI agent with mathematical safety verification. Every action is checked against constitutional rules using neuro-symbolic gradient steering. Every decision carries a cryptographic proof.

Not AI watching AI. Mathematics governing AI.

Quick Start

from evrguard import EvrmindGuard

guard = EvrmindGuard(api_url="https://evrguard.up.railway.app")

@guard.wrap_tool
def execute_query(sql):
    return db.execute(sql)

That is the entire integration. Three lines plus a decorator on each guarded tool.

Installation

pip install evrguard

Python 3.9 or newer. The only runtime dependency is requests (the SDK falls back to the standard library urllib if it is not present).

MCP Server for Claude Code

EvrGuard ships an MCP server. Add to your .mcp.json:

{
  "mcpServers": {
    "evrguard": {
      "command": "python",
      "args": ["-m", "evrguard.mcp_server"],
      "env": {
        "EVRGUARD_API_URL": "https://evrguard.up.railway.app"
      }
    }
  }
}

Restart Claude Code. The evrguard_verify tool is now available to the agent for every action.

Features

• Triple-layer detection — pattern matching, geometric safety net, and intent analysis. 97.3% hostile catch rate on the OWASP Agentic AI red-team baseline (87/87).
• Neuro-symbolic gradient steering — does not just block. Steers the agent toward a compliant alternative on a 32-dimensional constitutional manifold.
• Cryptographic proofs — every verdict carries an HMAC-SHA256 signature and a Merkle-anchored proof hash. Verifiable by any auditor.
• 10 industry constitutions — pre-built rule sets for finance, healthcare, legal, devops, retail, education, insurance, government, critical infrastructure, and general enterprise.
• OWASP Agentic Top 10 — preset compliance pack covering all ten agentic AI threats.
• Monitor mode — log verdicts without blocking. Run for a week before enforcing.
• Allowlist mode — for maximum control, allow only explicitly permitted actions and block everything else.

Industry Constitutions

Industry	Rules	Use For
General Enterprise	12	Any business using AI agents
Financial Services	14	Banks, trading, fintech, payments
Healthcare	12	Hospitals, clinics, health tech
Legal	11	Law firms, legal tech, compliance
DevOps / Engineering	12	Software teams, coding agents
Retail / E-commerce	11	Online retailers, marketplaces
Education	10	Schools, universities, edtech
Insurance	12	Insurers, brokers, insurtechs
Government	12	Public sector, civil service
Critical Infrastructure	14	Energy, water, telecoms, transport

All ten templates ship in the constitutions/ directory.

Integration Examples

Framework	Example
LangChain	examples/langchain_integration.py
LangGraph	examples/langgraph_integration.py
AutoGen	examples/autogen_integration.py
CrewAI	examples/crewai_integration.py
Claude Code	examples/claude_code_setup.md
OpenClaw	examples/openclaw_setup.md

Hardware Enforcement

For environments requiring physical enforcement, the same constitutional rules compile to the Evrmind safety processor on dedicated silicon. 7.47 ns interdiction. Same constitution. Two layers.

Learn more at labs.evrmind.io.

How it Works

   action ─▶  Layer 1: Pattern Matcher  ─▶
              Layer 2: Geometric Safety Net  ─▶
              Layer 3: Intent Analysis  ─▶  verdict + proof

The verification engine runs on the hosted EvrGuard API. The Python SDK in this repository is the open-source integration layer.

Licence

MIT. See LICENCE.

Copyright © 2026 Evrmind Limited.

---

labs.evrmind.io · GitHub · PyPI