---

The Problem

AI agents are shipping to production with zero runtime security. They can leak API keys in prompts, expose customer PII in responses, call unauthorized tools, and execute dangerous commands — and nobody finds out until the incident report.

Model safety training alone doesn't prevent these failures. You need enforcement at runtime, at the boundaries where data actually moves.

The Solution

SafeAI is the runtime security layer that sits between your AI agents and the outside world. It intercepts every prompt, tool call, and model response — blocking secrets, redacting PII, enforcing policies, and logging everything to an immutable audit trail.

Two lines to get started. Zero changes to your existing code.

from safeai import SafeAI

ai = SafeAI.quickstart()

# Block secrets before they reach any LLM
scan = ai.scan_input("Summarize this: token=sk-ABCDEF1234567890ABCDEF")
# => BLOCKED — secret detected, never reaches the model

# Redact PII from model responses before your users see them
guard = ai.guard_output("Contact alice@example.com or call 555-123-4567")
print(guard.safe_output)
# => "Contact [REDACTED] or call [REDACTED]"

Works with OpenAI, Anthropic, Google, LangChain, CrewAI, AutoGen, Claude Code, Cursor, and any HTTP-based AI stack.

How It Works

SafeAI enforces security at three runtime boundaries — the exact points where incidents happen:

                    ┌──────────────────────────────────┐
 User / Agent  ───▶ │  INPUT BOUNDARY   (scan_input)   │ ───▶ AI Provider
                    │  ACTION BOUNDARY  (intercept)    │      (OpenAI, Gemini,
 AI Provider   ◀─── │  OUTPUT BOUNDARY  (guard_output) │ ◀───  Claude, etc.)
                    └──────────────────────────────────┘
                               SafeAI Runtime

Boundary	What it stops
Input	Secrets, credentials, and sensitive data in prompts before they reach the model
Action	Unauthorized tool calls, dangerous commands (rm -rf /, DROP TABLE), and cross-agent messaging violations
Output	PII, internal data, and unsafe content in model responses before they reach users

Every decision is logged. Every action is auditable. No exceptions.

Quick Start

Install

pip install safeai-sdk

With extras for your stack:

pip install "safeai-sdk[vault]"   # HashiCorp Vault backend
pip install "safeai-sdk[aws]"     # AWS Secrets Manager backend
pip install "safeai-sdk[mcp]"     # MCP server for coding agents
pip install "safeai-sdk[all]"     # Everything

Scaffold a full project

safeai init --path .

This generates safeai.yaml, policies, contracts, agent identities, plugins, and alert rules — ready for production. Then:

ai = SafeAI.from_config("safeai.yaml")

Deploy as a sidecar proxy

Protect any AI stack without changing a single line of application code:

safeai serve --mode sidecar --host 127.0.0.1 --port 8910 --config safeai.yaml

curl http://127.0.0.1:8910/v1/health    # Health check
curl -s -X POST http://127.0.0.1:8910/v1/scan/input \
  -H "content-type: application/json" \
  -d '{"text":"hello world","agent_id":"default-agent"}'

AI-powered auto-configuration

Let SafeAI's intelligence layer analyze your codebase and generate tailored security policies:

safeai intelligence auto-config --apply

What SafeAI Does

🔒 Detection & Prevention

Capability	Description
Secret Detection	Blocks API keys, tokens, and credentials before they reach any LLM
PII Protection	Auto-redacts emails, phone numbers, SSNs, credit cards, and more
Content Moderation	80+ detectors for toxicity, prompt injection, jailbreaks — all local, no external APIs
Dangerous Commands	Stops rm -rf /, DROP TABLE, fork bombs, pipe-to-shell, force pushes

🛡️ Policy & Control

Capability	Description
Policy Engine	Priority-based YAML rules with tag hierarchies, hot reload, and boundary-specific matching
Tool Contracts	Declare what each tool can accept/emit — undeclared tools are denied by default
Agent Identity	Bind agents to specific tools and clearance levels for zero-trust enforcement
Approval Workflows	Human-in-the-loop gates for high-risk operations with TTL and dedup

🏗️ Infrastructure

Capability	Description
Encrypted Memory	Schema-enforced agent memory with field-level encryption and auto-expiry
Capability Tokens	Scoped, time-limited tokens for secret access — agents never see raw credentials
Audit Trail	Append-only logs with context hashes, filterable by agent, tool, session, and time
Multi-Provider Routing	Priority, cost, latency, and round-robin strategies with circuit-breaker failover

🚀 Operations & Scale

Capability	Description
Dashboard	Web UI for incidents, approvals, compliance summaries, and tenant/RBAC controls
Alerting	File, webhook, Slack, email, PagerDuty, and Opsgenie channels with Prometheus metrics
Skills System	Install pre-built packages — GDPR, HIPAA, PCI-DSS, prompt injection shield — in one command
Intelligence Layer	5 AI advisory agents for auto-config, policy recs, incident explanation, and compliance mapping
Cost Governance	Track token usage, enforce budgets with hard blocks, extract costs from provider responses

Works With Everything

AI Providers	Agent Frameworks	Coding Agents	Deployment
OpenAI	LangChain	Claude Code	Python SDK
Anthropic Claude	CrewAI	Cursor	REST API (sidecar)
Google Gemini	AutoGen	Copilot	Gateway proxy
Ollama	Claude ADK	Any MCP client	MCP server
Any HTTP API	Google ADK		CLI hooks

# Example: Wrap any LangChain tool in 3 lines
from safeai import SafeAI
from safeai.middleware.langchain import wrap_langchain_tool

ai = SafeAI.from_config("safeai.yaml")
safe_tool = wrap_langchain_tool(ai, my_tool, agent_id="default-agent")

CLI at a Glance

safeai init --path .                          # Scaffold config and policies
safeai serve --mode sidecar --port 8910       # Start the proxy server
safeai setup claude-code --config safeai.yaml # Install into Claude Code
safeai setup cursor --config safeai.yaml      # Install into Cursor
safeai mcp --config safeai.yaml               # Start MCP server
safeai intelligence auto-config --path .      # AI-powered auto-configuration
safeai skills add prompt-injection-shield     # Install a skill package
safeai scan --boundary input --input "text"   # Scan from the command line
safeai approvals list --status pending        # Manage approval queues
safeai observe agents                         # Agent observability

[!TIP] Run safeai --help or see the full CLI reference for all commands.

Extend with Plugins

Drop custom detectors, adapters, and policy templates into plugins/:

def safeai_detectors():
    return [(r"my-pattern", "custom.tag", "my_detector")]

def safeai_policy_templates():
    return [{"name": "my-template", "template": {"version": "v1alpha1", "policies": []}}]

Built-in template packs: finance, healthcare, support.

Documentation

Resource	Link
Getting started	Installation · Quickstart · Configuration
Guides	Policy Engine · Tool Contracts · Agent Identity · Audit Logging
Integrations	LangChain · CrewAI · AutoGen · Coding Agents
Reference	CLI · API · Architecture

Local Development

git clone https://github.com/enendufrankc/safeai.git
cd safeai
pip install -e ".[dev,all]"

Run the quality gates:

ruff check safeai tests        # Lint
mypy safeai                    # Type check
python -m pytest tests/ -v     # Test