Skip to content

Integrations Overview

SafeAI is framework-agnostic -- it works with any AI agent framework, coding assistant, or deployment topology. Choose the integration mode that fits your stack.

Auto-generate integration code

The intelligence layer can generate framework-specific integration code for your project: 

safeai intelligence integrate --target <framework> --path . --apply
See the Intelligence Layer guide for details.

Integration Modes

graph TB
    subgraph "Your AI Agents"
        LC[LangChain Agent]
        CR[CrewAI Agent]
        AG[AutoGen Agent]
        CA[Claude ADK Agent]
        GA[Google ADK Agent]
        CC[Claude Code]
        CU[Cursor]
        CUSTOM[Custom Agent]
    end

    subgraph "SafeAI Runtime Layer"
        direction TB
        SDK["SDK Adapters<br/>(in-process)"]
        HOOK["Coding Agent Hooks<br/>(stdin/stdout)"]
        PROXY["Proxy / Sidecar / Gateway<br/>(HTTP REST)"]
        PLUGIN["Plugin System<br/>(extensible)"]
    end

    subgraph "Enforcement"
        PE[Policy Engine]
        SD[Secret Detection]
        PII[PII Protection]
        TC[Tool Contracts]
        AL[Audit Log]
    end

    LC --> SDK
    CR --> SDK
    AG --> SDK
    CA --> SDK
    GA --> SDK
    CC --> HOOK
    CU --> HOOK
    CUSTOM --> PROXY

    SDK --> PE
    HOOK --> PE
    PROXY --> PE
    PLUGIN --> PE

    PE --> SD
    PE --> PII
    PE --> TC
    PE --> AL

SDK Adapters (In-Process)

Embed SafeAI directly inside your agent process. Each adapter wraps your framework's tool-calling interface so every invocation passes through SafeAI's policy engine.

Framework Adapter Page
LangChain ai.langchain_adapter() LangChain
CrewAI ai.crewai_adapter() CrewAI
AutoGen ai.autogen_adapter() AutoGen
Claude ADK ai.claude_adk_adapter() Claude ADK
Google ADK ai.google_adk_adapter() Google ADK 
from safeai import SafeAI

ai = SafeAI()
adapter = ai.langchain_adapter()          # or crewai_adapter(), autogen_adapter(), etc.
safe_tool = adapter.wrap_tool("my_tool", tool_fn, agent_id="agent-1")

3 lines to integrate

Every SDK adapter follows the same pattern: create SafeAI -> get adapter -> wrap tools. Your existing framework code stays unchanged.

Coding Agent Hooks

For AI-powered coding assistants that support shell hooks or MCP, SafeAI provides first-class setup commands.

Agent Command Page
Claude Code safeai setup claude-code Coding Agents
Cursor safeai setup cursor Coding Agents
MCP clients safeai mcp Coding Agents
Any agent safeai hook Coding Agents 
# One command to protect your coding agent
safeai setup claude-code

How hooks work

The safeai hook command reads a JSON action from stdin, evaluates it against the active policy, and writes a JSON decision to stdout. This simple protocol works with any agent that supports shell-based tool approval.

Proxy / Sidecar / Gateway

Run SafeAI as a standalone HTTP service. Perfect for polyglot environments, multi-agent deployments, or when you cannot modify agent source code.

Mode Command Use Case
Sidecar safeai serve --mode sidecar --port 8000 Single-agent, same host
Gateway safeai serve --mode gateway Multi-agent, centralized enforcement 
# Start a sidecar next to your agent
safeai serve --mode sidecar --port 8000

# Scan input from any language
curl -X POST http://localhost:8000/v1/scan/input \
  -H "Content-Type: application/json" \
  -d '{"text": "Process this request"}'

See Proxy / Sidecar for full endpoint reference.

Plugin System

Extend SafeAI with custom detectors, adapters, and policy templates -- no forking required.

# plugins/my_plugin.py
def safeai_detectors():
    return [MyCustomDetector()]

def safeai_adapters():
    return {"my_adapter": MyAdapter}

# safeai.yaml
plugins:
  enabled: true
  plugin_files:
    - "plugins/*.py"

See Plugins for the full plugin API.

Choosing the Right Mode

flowchart TD
    A["Can you modify<br/>agent source code?"] -->|Yes| B["Which framework?"]
    A -->|No| C["Proxy / Sidecar"]
    B -->|LangChain / CrewAI / AutoGen| D["SDK Adapter"]
    B -->|Claude ADK / Google ADK| D
    B -->|Claude Code / Cursor| E["Coding Agent Hook"]
    B -->|Custom / Other| F["SDK + Plugin<br/>or Proxy"]
Criteria SDK Adapter Coding Hook Proxy/Sidecar
Latency Lowest (in-process) Low (subprocess) Medium (HTTP)
Language support Python only Any (JSON protocol) Any (HTTP)
Source code changes Required None None
Multi-agent Per-agent adapter Per-agent hook Centralized gateway
Metrics In-process CLI output Prometheus endpoint

Next Steps