SupraWall

The rule of law for zero-human companies.

A deterministic security perimeter for AI agents. One line of code. Open source.

Quickstart · How it works · Frameworks · EU AI Act templates · Cloud · Docs

Paperclip gives your agents a company. SupraWall gives them a constitution.

---

Why this exists

AI agents now write code, spend money, query databases, and take real-world actions on your behalf — autonomously. The frameworks that orchestrate them (Paperclip, OpenClaw, LangChain, CrewAI, AutoGen, Claude Code) are excellent at making them productive. None of them are responsible for making them safe.

So agents do what unconstrained software has always done: leak credentials, run DROP TABLE users, exfiltrate PII, burn $40k overnight in OpenAI tokens, and fail every compliance audit you'll ever face under the EU AI Act.

SupraWall is a deterministic perimeter that wraps your agent — any agent — and intercepts every tool call before it executes. Not probabilistically. Not via another LLM. Not after the fact. At the boundary, in under 2ms, with a signed audit log.

It is not another guardrail model. It is the rule of law.

And with the EU AI Act enforcement deadline on August 2, 2026, we ship 8 pre-built sector templates covering every Annex III high-risk category — HR, healthcare, education, critical infrastructure, biometrics, law enforcement, migration, justice — plus a DORA template for financial services. Compliance by pip install, not by 200-page PDF.

---

Quickstart

Python

pip install suprawall-sdk

from suprawall import secure_agent
from your_framework import build_agent  # LangChain, CrewAI, AutoGen, custom — doesn't matter

agent = secure_agent(build_agent(), policy="policies/langchain-safe.json")

# Dangerous tool call from a prompt-injected user message:
agent.invoke("DROP TABLE users")
# ⚡ SupraWall intercepted. BLOCKED. Audit log #A-00847 signed ✓

TypeScript / Node

npm install suprawall

import { secureAgent } from "suprawall";
import { agent } from "./my-agent";

const safe = secureAgent(agent, { policy: "./policies/langchain-safe.json" });
await safe.invoke("DROP TABLE users"); // BLOCKED — pre-execution

That's it. No proxy to deploy. No sidecar. No model fine-tune. The wrapper sits between the agent's reasoning loop and the tool runtime, where deterministic rules belong.

---

How it works

Three layers, evaluated in order. Local policy always wins.

#	Layer	Latency	What it does
1	Pre-Execution Interception	<1ms	Every tool call routed through SupraWall before the runtime sees it. Hard-coded — no LLM in the loop.
2	Zero-Trust Policy Enforcement	<2ms	Budget caps, PII scrubbing, SQL/shell injection blocks, credential vault, allow/deny lists — enforced as code, not as suggestions.
3	Compliance Audit Trail	async	Every decision RSA-signed, timestamped, exportable. Maps to EU AI Act Art. 9, 13, 14 out of the box.

The semantic AI layer (Layer 2.5, optional, cloud-only) catches context-dependent attacks that regex can't see — but local deterministic policy is always the first and final word.

---

Works with any agent stack

SupraWall is framework-agnostic. It wraps the tool boundary, which every agent has.

Framework	Status	Plugin
Paperclip	✅ First-class	packages/paperclip-plugin
LangChain (Py + TS)	✅ First-class	Built into core SDK
CrewAI	✅ First-class	Built into core SDK
AutoGen	✅ First-class	Built into core SDK
Claude Code / OpenClaw	✅ Via MCP	suprawall-mcp-plugin
Vercel AI SDK	✅ First-class	Built into core SDK
Custom / homegrown	✅	One-line secure_agent() wrapper

Languages: Python, TypeScript, Go, C#. More via the MCP plugin.

---

What it stops

Threat	How SupraWall stops it	EU AI Act
Credential theft	Vault injects secrets at runtime. Agents never see real keys. Logs scrubbed in 5+ encodings.	Art. 13
Runaway costs	Hard per-agent budget caps, per-model token accounting, circuit breakers.	Art. 9
Unauthorized actions	Deterministic ALLOW/DENY policies block tool calls before execution.	Art. 9
PII exposure	Response scrubbing redacts SSN, CC, email, custom regex — across encodings.	Art. 13
No audit trail	RSA-signed logs with risk scores. Exportable as compliance evidence.	Art. 13
No human oversight	REQUIRE_APPROVAL pauses the agent and notifies a human before high-risk actions.	Art. 14
Prompt-injection-driven actions	Local policy ignores agent intent — only the tool-call signature matters.	Art. 9

---

Built-in policy templates

npx suprawall init  # interactive policy bootstrap

Policy	Protects against
langchain-safe.json	rm -rf, .env reads, unwhitelisted shell
pii-protection.json	SSN, CC, email exfiltration
eu-ai-act-audit.json	Human-in-the-loop for high-risk tools
budget-guardrail.json	Token + cost circuit breakers
paperclip-company.json	Company-scoped budgets, role-based tool access

→ All starter policies · → Write your own

---

EU AI Act compliance — shipped, not promised

Enforcement begins August 2, 2026. Every Annex III high-risk sector needs a documented, enforceable risk management system (Art. 9), a tamper-proof audit trail (Art. 13), and human oversight (Art. 14). Most teams are going to scramble. You don't have to.

SupraWall ships 8 pre-built sector templates covering every Annex III high-risk category — plus a Banking & Finance template mapped to DORA. Each one is a real enforcement config with DENY rules, REQUIRE_APPROVAL gates, mandatory logging, and a conformity-assessment path built in.

Sector	Annex III	Risk level	Conformity	What it blocks out of the box
Biometrics	Category 1	Critical	Third-party	Real-time ID in public spaces, emotion recognition without approval
Critical Infrastructure	Category 2	Critical	Self	Physical-action tools without human confirm, unsafe disconnection
Education	Category 3	High	Self	Autonomous admission rejections, scoring without explainability
HR & Employment	Category 4	High	Self	Autonomous hire/fire, salary changes, performance reviews without sign-off
Healthcare	Category 5	Critical	Third-party	Diagnosis without human review, PHI exfiltration, unlogged patient actions
Law Enforcement	Category 6	Critical	Third-party	Predictive policing outputs without review, autonomous evidence decisions
Migration & Border	Category 7	High	Self	Automated visa denials, risk-scoring without human
Justice & Democracy	Category 8	High	Self	Autonomous judicial outputs, election-related agent actions
Banking & Finance (DORA)	—	High	Self	Autonomous trading, unlogged client-facing decisions

Apply a sector template in one line:

from suprawall import secure_agent
agent = secure_agent(build_agent(), template="hr-employment")

Every template includes the baseline controls every Annex III system needs (risk management log, data-quality gate, human oversight hook, post-market monitoring, incident reporting) and layers sector-specific overrides on top. Every policy decision is RSA-signed and exportable as compliance evidence — the kind your auditor will actually accept.

Why this matters right now: August 2, 2026 is months away. The penalty for non-compliance is up to €35M or 7% of global turnover. If your auditor asks "what stopped the agent from terminating an employee autonomously?" — you hand them signed log entry #A-00847. Most teams will be hand-waving. You'll be compliant by pip install.

→ Full EU AI Act compliance guide · → Sector templates source

---

Self-host or cloud

SupraWall is fully open source under Apache 2.0 — clone it, run it, ship it.

	Open Source (Self-Hosted)	Cloud
Layer 1 deterministic engine	✅ Free forever	✅
All built-in policies	✅ Free forever	✅
RSA-signed audit log	✅ Free forever	✅
Layer 2.5 semantic AI detection	—	✅
Hosted dashboard + multi-tenant	—	✅
Compliance report generation	—	✅
SLA + support	—	✅

# Self-host the dashboard
docker compose up

→ Deploy on cloud · → AWS Marketplace

---

Why "deterministic" matters

The dominant security pattern for agents today is another LLM judging the first LLM (guardrail models, classifier filters, etc.). That works ~80% of the time, fails silently the other 20%, costs tokens on every call, and produces unauditable decisions.

SupraWall takes the opposite stance: rules belong in code, not in prompts. A deterministic policy either matches or it doesn't. The decision is reproducible, the latency is constant, the audit trail is real, and there's no prompt you can write to talk it out of doing its job.

If you want probabilistic content moderation, use a guardrail model. If you want to stop your agent from wiring funds to the wrong account, use a deterministic perimeter.

---

Star history

If SupraWall saved you from an incident, please ⭐ the repo — it's how this kind of infrastructure finds the people who need it.

---

Contributing

We're a small team (Wiser Automation) and we want SupraWall to be a community-owned standard, not a single-vendor tool.

• Contributing guide
• Code of Conduct
• Security policy — please report vulnerabilities privately
• Roadmap

Active issues good for first-time contributors are tagged good first issue.

---

Links

Website · Docs · Cloud · Blog · X / @The_real_Peghin · License: Apache 2.0

_{Built by Wiser Automation · Made for the zero-human company era.}