Skip to content

integration-patterns.md

Latest commit

 

History

History
166 lines (109 loc) · 5.54 KB

integration-patterns.md

File metadata and controls

166 lines (109 loc) · 5.54 KB
Raw

Integration Patterns

Practical reference for how different agent frameworks and HTTP clients work inside SILO's sandboxed environment.

Overview

SILO makes external agents work through two complementary mechanisms:

  1. http_proxy / https_proxy env vars — automatically set inside the container. Most HTTP clients route through these without any code changes.
  2. llm.gateway DNS alias — resolves to the host machine, letting agents target http://llm.gateway:8199 as their LLM endpoint.

This document covers how specific frameworks and libraries integrate with each mechanism.

Python Agents

requests / httpx / urllib3 / aiohttp

These libraries automatically respect the http_proxy and https_proxy environment variables set inside the container. No code changes needed — your existing code routes through the SILO proxy transparently.

For aiohttp, pass trust_env=True to ClientSession.

OpenAI Python SDK

Override the base URL to point at the proxy:

from openai import OpenAI

client = OpenAI(
    base_url="http://llm.gateway:8199/v1",
    api_key="sk-placeholder",  # Required by SDK, but proxy replaces it
)

response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello from the sandbox!"}],
)

Anthropic Python SDK

from anthropic import Anthropic

client = Anthropic(
    base_url="http://llm.gateway:8199",
    api_key="sk-ant-placeholder",  # Proxy injects the real key
)

response = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello from the sandbox!"}],
)

Node.js Agents

Node's built-in fetch does not automatically use proxy env vars. Libraries like axios do. For the simplest path, use the OpenAI Node SDK with a base URL override:

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "http://llm.gateway:8199/v1",
  apiKey: "sk-placeholder", // Proxy injects the real key
});

const response = await client.chat.completions.create({
  model: "gpt-4o",
  messages: [{ role: "user", content: "Hello from the sandbox!" }],
});

For raw fetch, point directly at the proxy: fetch("http://llm.gateway:8199/v1/chat/completions", ...).

LangChain / LangGraph

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="gpt-4o",
    base_url="http://llm.gateway:8199/v1",
    api_key="sk-placeholder",
)

response = llm.invoke("What is the capital of France?")

For Anthropic, use ChatAnthropic with anthropic_api_url="http://llm.gateway:8199". For LangGraph agents, tools execute inside the sandbox — fully isolated.

Set LANGCHAIN_TRACING_V2=false to disable external tracing (it would be blocked by policy anyway).

Ollama / Local LLMs

During silo onboard, choose Local as the provider and set Base URL to your local server (e.g. http://localhost:11434). The proxy translates llm.gateway requests to your local server automatically:

Container → http://llm.gateway:8199/v1/chat/completions
         → Proxy forwards to → http://localhost:11434/v1/chat/completions

Agent code uses the same base_url="http://llm.gateway:8199/v1" pattern — no changes needed.

Agents That Browse the Web

Rather than allowing all internet access, whitelist specific domains:

silo policy add "*.google.com" --comment "Google search"
silo policy add "*.wikipedia.org" --comment "Wikipedia"
silo policy add "*.github.com" --comment "GitHub"

Alternatively, start with Strict mode (chosen during silo onboard) and use the interactive interceptor — each unauthorized request triggers a real-time approve/deny prompt on the host terminal. See Running External Agents for details.

Multi-Agent on Shared Network

When running multiple agents with silo start --name, all containers share the silo-sandbox-net bridge network. Containers can reach each other by name (e.g. curl http://silo-backend:9000). All agents share the same egress policy but have isolated /workspace mounts. See Multi-Agent Support for full details.

Common Environment Variable Patterns

Quick reference for environment variables to set inside the container for different frameworks:

Framework Variables
OpenAI SDK OPENAI_BASE_URL=http://llm.gateway:8199/v1 OPENAI_API_KEY=sk-placeholder
Anthropic SDK ANTHROPIC_BASE_URL=http://llm.gateway:8199 ANTHROPIC_API_KEY=sk-ant-placeholder
LangChain OPENAI_API_BASE=http://llm.gateway:8199/v1 OPENAI_API_KEY=sk-placeholder
curl Uses http_proxy/https_proxy automatically (already set)
pip Uses https_proxy automatically (already set)
npm Uses https_proxy automatically (already set)
git Uses http_proxy/https_proxy automatically (already set)

Tips for Agent Authors

If you're building an agent that should work well inside SILO:

  1. Respect proxy env vars — Use HTTP libraries that honor http_proxy / https_proxy
  2. Make the base URL configurable — Allow users to override the LLM endpoint via env var
  3. Don't hardcode API keys — Accept them via environment variable, not config files
  4. Handle HTTP 403 gracefully — SILO returns 403 for blocked requests; don't crash on policy denials
  5. Minimize required network access — The fewer hosts your agent needs, the easier it is to sandbox

SILO v0.2.0 — Secure AI Agent Sandbox Orchestrator