惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

Simon Willison's Weblog
Simon Willison's Weblog
T
Threatpost
Threat Intelligence Blog | Flashpoint
Threat Intelligence Blog | Flashpoint
Spread Privacy
Spread Privacy
cs.AI updates on arXiv.org
cs.AI updates on arXiv.org
cs.CV updates on arXiv.org
cs.CV updates on arXiv.org
Engineering at Meta
Engineering at Meta
T
Tenable Blog
C
Cisco Blogs
T
The Blog of Author Tim Ferriss
NISL@THU
NISL@THU
T
Threat Research - Cisco Blogs
T
The Exploit Database - CXSecurity.com
P
Privacy & Cybersecurity Law Blog
Recent Commits to openclaw:main
Recent Commits to openclaw:main
S
Secure Thoughts
N
News and Events Feed by Topic
Google DeepMind News
Google DeepMind News
博客园 - 三生石上(FineUI控件)
Microsoft Azure Blog
Microsoft Azure Blog
月光博客
月光博客
H
Hacker News: Front Page
I
InfoQ
L
LangChain Blog
Security Latest
Security Latest
The Cloudflare Blog
Forbes - Security
Forbes - Security
D
Darknet – Hacking Tools, Hacker News & Cyber Security
Stack Overflow Blog
Stack Overflow Blog
TaoSecurity Blog
TaoSecurity Blog
量子位
Cyber Security Advisories - MS-ISAC
Cyber Security Advisories - MS-ISAC
V
Visual Studio Blog
Scott Helme
Scott Helme
爱范儿
爱范儿
A
Arctic Wolf
F
Full Disclosure
酷 壳 – CoolShell
酷 壳 – CoolShell
Schneier on Security
Schneier on Security
N
News and Events Feed by Topic
C
CXSECURITY Database RSS Feed - CXSecurity.com
L
LINUX DO - 最新话题
V2EX - 技术
V2EX - 技术
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
人人都是产品经理
人人都是产品经理
The Hacker News
The Hacker News
Hugging Face - Blog
Hugging Face - Blog
Security Archives - TechRepublic
Security Archives - TechRepublic
Know Your Adversary
Know Your Adversary
Application and Cybersecurity Blog
Application and Cybersecurity Blog

DEV Community

Authentication Security Deep Dive: From Brute Force to Salted Hashing (With Java Examples) Why AI Systems Don’t Fail — They Drift Spilling beans for how i learn for exam😁"Reinforcement Learning Cheat Sheet" I Replaced Chrome with Safari for AI Browser Automation. Here's What Broke (and What Finally Worked) How Python Borrows Other People's Work The $40 Architecture: Processing 1 Billion API Requests with 99.99% Uptime Vibe Coding: A Workflow Guide (From Zero to SaaS) Most webhook security guides protect the wrong side. The scary part is delivery. Headless CMS for TanStack Start: Build a Blog with Cosmic EU Age Verification App "Hacked in 2 Minutes" — What Actually Happened Comfy Cloud’s delete function does not actually remove files Running AI Models on GPU Cloud Servers: A Beginner Guide Event-driven media intelligence with AWS Step Functions and Bedrock I scored 500 AI prompts across 8 quality dimensions — here's what broke How to Call Google Gemini API from Next.js (Free Tier, No Backend Needed) The Portal Protocol: Reclaiming Human Connection in the Age of AI How to Fix Your Team's Scattered Knowledge Problem With a Self-Hosted Forum Intro to tc Cloud Functors: A Graph-First Mental Model for the Modern Cloud Designing Multi-Tenant Backends With Both Ownership and Team Access I Built a Neumorphic CSS Library with 77+ Components — Here's What I Learned PostgreSQL Performance Optimization: Why Connection Pooling Is Critical at Scale Cómo construí un SaaS multi-rubro para gestionar expensas en Argentina con FastAPI + Vue 3 🚀 I Built an Ethical Hacking Scanner Tool – Open Source Project I Replaced /usage and /context in Claude Code With a Single Statusline A Pythonic Way to Handle Emails (IMAP/SMTP) with Auto-Discovery and AI-Ready Design I Collected 8.9 Million Polymarket Price Points — Here's What I Found About How Markets Really Move EcoTrack AI — Carbon Footprint Tracker & Dashboard Everyone's Using AI. No One Agrees How. 5 self-hosted ebook managers worth trying in 2026 Building Your First AI Agent with LangChain: From Chatbot to Autonomous Assistant Common SOC 2 Failures (Real World) Stop Vibe-Checking Your AI App: A Practical Guide to Evals How to Use SonarQube and SonarScanner Locally to Level Up Your Code Quality Your Next To-Do App Is Dead — I Replaced Mine with an OpenClaw AI Sign a Nostr event in 60 lines of Python using coincurve — no nostr-sdk, no nbxplorer, no rust toolchain ITGC Audit Explained Like You’re in Big 4 Patch Tuesday abril 2026: Microsoft parcha 163 vulnerabilidades y un zero-day en SharePoint Stop scraping everything: a better way to track competitor price changes Listing on MCPize + the Official MCP Registry while routing payments OUTSIDE the marketplace — how I kept 100% of my x402 revenue Building an AI-Powered Risk Intelligence System Using Serverless Architecture Why We Ripped Function Overloading Out of Our AI Toolchain Testing AI-Generated Code: How to Actually Know If It Works SaaS Churn Is Killing Your Business. Here Is What to Do About It (Without a Support Team) The Speed of AI Is No Longer Linear - And Self-Improving Models Are Why How to Implement RBAC for MCP Tools: A Practical Guide for Engineering Teams From Standard Quote to Persuasive Proposal: AI Automation for Arborists I built a CLI that scaffolds complete multi-tenant SaaS apps Axios CVE-2025–62718: The Silent SSRF Bug That Could Be Hiding in Your Node.js App Right Now The dashboard that ended our friendship Data Pipelines Explained Simply (and How to Build Them with Python) The Hidden Cost of AI Systems Nobody Talks About. undefined vs undeclared, and how typeof behaves Switching from file-based jobs to NATS/Kafka in Rust without changing code io_uring Adventures: Rust Servers That Love Syscalls Why Agentic AI is Killing the Traditional Database The POUR principles of web accessibility for developers and designers Quantum Neural Network 3D — A Deep Dive into Interactive WebGL Visualization How To Install Caveman In Codex On macOS And Windows Automation Pipeline Reliability: Why Your Workflow Breaks When Nobody Is Watching I Built an 'Open World' AI Coding Agent — It Works From ANY Folder From Freelancing to Product: A Tech Service Company's SaaS Transformation China's AI Giants: Adding Tencent Hunyuan & ByteDance Doubao to AI University (74 Providers) On the Vibe Coders and Their Lies clerk: Auto-Summarize Your Claude Code Sessions AI Weekly — 2026/04/10–04/17 | The Model Lockdown Is Here, but the Toolchain Is the Real Battleground AI 週報 — 2026/04/10–2026/04/17 模型封鎖潮來了,但工具鏈才是真戰場 Maybe this is how Open-Source apps are born... 🚀 Fine-Tune LLMs with LoRA and QLoRA: 2026 Guide tRPC v11 + Next.js App Router: End-to-End Type Safety Without the Boilerplate ShadCN UI in 2026: Why I Stopped Installing Component Libraries and Started Owning My Components SaaS Billing in React Server Components: Stripe + Supabase Without a Single `useEffect` Join our DEV Weekend Challenge — $1,000 in Prizes Across TEN winners! Submissions Due April 20 at 6:59 AM UTC. Implementing FSRS Spaced Repetition in Flutter + Supabase — Adding Memory Science to an AI Learning App "I Texted My Localhost From the Train — Claude Code Fixed the Bug Before I Got Home" I Built a Sales Prep AI and It Went Deeper Than Expected Design to Code #2: One JSON, Eleven Outputs Solving the 100M-Row Problem: A Summary Table Pattern for High-Volume Push Notification Logs Flutter Web With Wasm: What Actually Changes For Developers I Built 50 Royalty-Free Soundtracks for My Side Project in a Weekend Using AI Music Generation The Vibe Coding Security Checklist: 7 Things to Check Before You Ship Stop Letting Googlebot Guess Fix Your React App's SEO Right Desconstruindo o Streaming do LinkedIn: Como Criar um Engine de Extração de Vídeo de Alta Performance com HLS e FFmpeg (EDA Part-1) EDA (Exploratory Data Analysis) Explained With Real Life — Why Looking at Your Data Is the Most Important Step in Machine Learning Brand Relationship Management at Scale: Our 4-Touch Outreach System for 200+ Brands Why String.fromEnvironment() Might Return an Empty String in Dart JGuardrails 1.0.0 — Hardening Java LLM Apps Against Jailbreaks, Toxicity, and Prompt Injection Plan and Schedule a Full Week of Threads Content From One Claude Conversation Coding Cat Oran Ep3, Five Tables Changed Everything Updated: BFF Pattern I'm done watching freelancers get buried by 200 proposals. So I'm building the alternative. This is my first post BFS Algorithm in Java Step by Step Tutorial with Examples Tracking LLM Pricing Monthly: An Open Dataset for 22 AI Models How We Measure Content ROI on a Comparison Site: Revenue Attribution Without Perfect Data Introducing Nova AI Ops: The AI-Native Operating System for SRE Teams I built a free desktop video downloader for Windows — Grabbit How Talkie OCR Helps Vision-Impaired & Dyslexic Users Read the World Around Them VRCFaceTracking安装和iPhone面捕配置教程,有bug Even CrowdStrike Can't See Your Agents The Automation Gold Rush: What n8n Workflows and Claude Are Opening Up for Developers Right Now
Hyperlight Sandbox + MCP CodeMode: Secure Agent Execution Beyond Containers
Thang Chung · 2026-05-11 · via DEV Community

Introduction

Most MCP CodeMode implementations today still execute generated code inside local processes, shared-kernel containers, or userspace sandboxes that ultimately rely on the host Linux kernel as the primary security boundary. In the previous post, OpenSandbox already demonstrated how production-grade sandbox orchestration can significantly improve safety compared to running generated Python directly on the host machine. But once agents start executing untrusted code, third-party tools, long-running workflows, or multi-tenant workloads, the challenge shifts from simple execution isolation to containment and blast-radius reduction.

Platforms like OpenSandbox solve a large part of this operational problem extremely well. OpenSandbox provides orchestration, sandbox lifecycle management, SDK abstraction, browser/code execution, and production-ready runtime operations for agent systems. In practice, however, the execution layer still typically depends on Docker containers and external sandbox orchestration. That introduces additional runtime layers, container lifecycle overhead, and more boundary transitions between the MCP runtime and the isolated execution environment.

This is where Hyperlight Sandbox becomes interesting. Instead of treating sandboxing primarily as infrastructure orchestration, Hyperlight pushes isolation directly into the runtime layer itself: hypervisor-backed microVM execution, no full Linux guest OS, reduced attack surface, fast startup, and WASI-friendly capability boundaries. Combined with MCP CodeMode, this creates a very different execution model for AI agents — one that moves closer to microVM-grade isolation without the operational and resource overhead of traditional virtual machines.

I submitted a feature request at Hyperlight Sandbox Issue #13. Just a couple of days later, the Hyperlight team added .NET SDK support in Hyperlight Sandbox PR #46.

I then spent a weekend experimenting with this approach, and in this blog post. I will walk through step by step what I have done so far.

Let's get started...

The MCP CodeMode approach with Hyperlight sandbox

MCP Codemode architecture

Ref: Mermaid

As you can see, the Hyperlight sandbox is an in-process call, meaning it runs inside a hypervisor-backed runtime embedded in the host application (look at the left box in the picture above). As a result, the execution experience feels like a native function call.

Step by step to set up the code

First of all, you have to reference some NuGet package for the Hyperlight sandbox at https://www.nuget.org/packages?q=HyperlightSandbox

<PackageReference Include="Hyperlight.HyperlightSandbox.Api" Version="0.4.0" />
<PackageReference Include="Hyperlight.HyperlightSandbox.Extensions.AI" Version="0.4.0" />
<PackageReference Include="Hyperlight.HyperlightSandbox.Guest.Python" Version="0.4.0" />

Enter fullscreen mode Exit fullscreen mode

In the previous part, we have already had the ISandboxRunner interface:

public interface ISandboxRunner
{
    string SyntaxGuide { get; }

    Task<RunnerResult> RunAsync(string code, CancellationToken ct);
}

Enter fullscreen mode Exit fullscreen mode

Now I just implemented this interface for the Hyperlight sandbox as below

public sealed class HyperlightSandboxRunner : ISandboxRunner, IDisposable
{
    private static readonly Regex AbsoluteHttpUrlRegex = new(
        "https?://[^\\s\"'\\)\\]]+",
        RegexOptions.IgnoreCase | RegexOptions.Compiled | RegexOptions.CultureInvariant);

    private static readonly ActivitySource ActivitySource = new("McpServer.CodeMode.HyperlightSandboxRunner");

    private readonly TimeSpan timeout;
    private readonly int maxToolCalls;
    private readonly ILogger<HyperlightSandboxRunner> logger;
    private readonly IReadOnlyList<Uri> allowedBaseUris;
    private readonly Sandbox sandbox;
    private bool disposed;

    public HyperlightSandboxRunner(
        TimeSpan timeout,
        int maxToolCalls,
        ILogger<HyperlightSandboxRunner> logger,
        IReadOnlyList<string>? allowedBaseUrls = null)
    {
        this.timeout = timeout;
        this.maxToolCalls = maxToolCalls;
        this.logger = logger;
        allowedBaseUris = NormalizeAllowedBaseUris(allowedBaseUrls);

        SandboxBuilder builder = new SandboxBuilder()
            .WithPythonModule()
            .WithTempOutput();

        sandbox = builder.Build();
        foreach (Uri allowedBaseUri in allowedBaseUris)
        {
            sandbox.AllowDomain(allowedBaseUri.AbsoluteUri.TrimEnd('/'));
        }
    }

    public string SyntaxGuide
    {
        get
        {
            string defaultBaseUrl = allowedBaseUris.Count > 0
                ? allowedBaseUris[0].AbsoluteUri.TrimEnd('/')
                : string.Empty;

            string baseUrlSection = defaultBaseUrl.Length > 0
                ? $"Use the injected `BASE_URL` variable for API calls. Default `BASE_URL` is \"{defaultBaseUrl}\"."
                : "Use the injected `BASE_URL` variable for API calls. No default URL was discovered from OpenAPI sources.";

            string allowedSection = allowedBaseUris.Count > 0
                ? $"Only call API URLs under configured OpenAPI bases via `BASE_URL`: {string.Join(", ", allowedBaseUris.Select(static u => $"\"{u.AbsoluteUri.TrimEnd('/')}\""))}."
                : "No API base allowlist is currently configured, so URL host restrictions are not enforced.";

            return $$"""
        Runner: hyperlight (Python sandbox)
        Write pure Python code.
        For HTTP calls in this sandbox, prefer built-in `http_get(url)` and `http_post(url, body="")`.
        A lightweight `requests` compatibility shim object is preloaded as `requests` for fallback compatibility.
        Do not import `requests` in generated code.
        Do NOT rely on `urllib`, `http.client`, or other stdlib HTTP modules in guest code.
        Prefer assigning the final value to `result`.
        If `result` is not set, stdout fallback is returned when available.
        {{baseUrlSection}}
        {{allowedSection}}
        Code mode is isolated from tool-Search tools.
        Do NOT use: SearchTools, CallTool, Search, GetSchema, or Execute in this code.
        Example:
            resp = http_get(f"{BASE_URL}/pet/findByStatus?status=sold")
            if resp["status"] >= 400:
                raise Exception(f"HTTP {resp['status']}")
            result = resp["body"]
        The final `result` value (or stdout fallback) is returned as tool output.
        """;
        }
    }

    public async Task<RunnerResult> RunAsync(string code, CancellationToken ct)
    {
        ThrowIfDisposed();
        ArgumentException.ThrowIfNullOrWhiteSpace(code);

        logger.LogInformation(
            "Hyperlight constrained execution requested. Timeout: {Timeout}, MaxToolCalls: {MaxToolCalls}.",
            timeout,
            maxToolCalls);

        using Activity? activity = ActivitySource.StartActivity("codemode.run", ActivityKind.Internal);
        activity?.SetTag("mcp.code.length", code.Length);
        activity?.SetTag("mcp.Execute.timeout.ms", timeout.TotalMilliseconds);
        activity?.SetTag("mcp.Execute.maxToolCalls", maxToolCalls);

        if (SandboxCodeGuard.ContainsForbiddenMetaToolUsage(code))
        {
            throw new InvalidOperationException(
                "Code mode is isolated from tool-Search tools. " +
                "Do not call SearchTools, CallTool, Search, GetSchema, or Execute inside code mode; use pure Python compute only. " +
                "If you need an MCP tool, call it outside Execute. If you stay in Execute, rewrite the task as direct Python logic or HTTP requests with http_get/http_post.");
        }

        if (ContainsForbiddenHardcodedApiUsage(code))
        {
            throw new InvalidOperationException(
                "Code mode HTTP calls must stay within configured OpenAPI data sources. " +
                "Use BASE_URL (or ALLOWED_BASE_URLS) instead of hardcoded external URLs.");
        }

        logger.LogInformation("Generated Python code for Hyperlight Execute:\n{Code}", code);

        using CancellationTokenSource timeoutCts = new(timeout);
        using CancellationTokenSource linkedCts = CancellationTokenSource.CreateLinkedTokenSource(ct, timeoutCts.Token);

        ExecutionResult executionResult = await Task.Run(() => ExecuteInSandbox(code), linkedCts.Token);

        activity?.SetTag("mcp.Execute.callCount", 0);
        activity?.SetTag("mcp.Execute.hasFinalValue", executionResult.Success);

        object? finalValue = ExtractFinalValue(executionResult);
        return new RunnerResult(finalValue, 0);
    }

    private ExecutionResult ExecuteInSandbox(string code)
    {
        string serializedCode = JsonSerializer.Serialize(code);
        string serializedDefaultBaseUrl = JsonSerializer.Serialize(
            allowedBaseUris.Count > 0 ? allowedBaseUris[0].AbsoluteUri.TrimEnd('/') : string.Empty);

        string wrapper = $$"""
code = {{serializedCode}}
BASE_URL = {{serializedDefaultBaseUrl}}

try:
    import sys
except Exception:
    sys = None

_native_http_get = None
_native_http_post = None

try:
    _native_http_get = http_get
except Exception:
    _native_http_get = None

try:
    _native_http_post = http_post
except Exception:
    _native_http_post = None

if _native_http_get is None or _native_http_post is None:
    try:
        from hyperlight import http_get as _hl_http_get, http_post as _hl_http_post
        if _native_http_get is None:
            _native_http_get = _hl_http_get
        if _native_http_post is None:
            _native_http_post = _hl_http_post
    except Exception:
        pass

def http_get(url):
    if _native_http_get is None:
        raise Exception("http_get is unavailable in this Hyperlight guest runtime")
    return _native_http_get(url)

def http_post(url, body=""):
    if _native_http_post is None:
        raise Exception("http_post is unavailable in this Hyperlight guest runtime")
    return _native_http_post(url, body)

def _encode_query(params):
    if params is None:
        return ""
    if isinstance(params, dict):
        parts = []
        for key, value in params.items():
            if isinstance(value, (list, tuple)):
                for item in value:
                    parts.append(str(key) + "=" + str(item))
            else:
                parts.append(str(key) + "=" + str(value))
        return "&".join(parts)
    return str(params)

def _append_query(url, params):
    query = _encode_query(params)
    if not query:
        return url
    return url + ("&" if "?" in url else "?") + query

class _RequestsHttpError(Exception):
    pass

class _RequestsResponse:
    def __init__(self, status_code, text):
        self.status_code = int(status_code) if status_code is not None else 0
        self.text = text if text is not None else ""
        self.content = self.text.encode("utf-8") if hasattr(self.text, "encode") else self.text

    def json(self):
        try:
            import json as _json
            return _json.loads(self.text)
        except Exception:
            return self.text

    def raise_for_status(self):
        if self.status_code >= 400:
            raise _RequestsHttpError("HTTP " + str(self.status_code))

def _requests_request(method, url, params=None, data=None, json=None, headers=None, timeout=None):
    del headers, timeout
    method_upper = str(method).upper()
    target = _append_query(str(url), params)

    if method_upper == "GET":
        raw = http_get(target)
    elif method_upper == "POST":
        body = data
        if json is not None:
            try:
                import json as _json
                body = _json.dumps(json)
            except Exception:
                body = str(json)
        if body is None:
            body = ""
        raw = http_post(target, str(body))
    else:
        raise Exception("Method not supported in hyperlight requests shim: " + method_upper)

    if isinstance(raw, dict):
        return _RequestsResponse(raw.get("status", 0), raw.get("body", ""))
    return _RequestsResponse(0, str(raw))

class _RequestsModule:
    RequestException = Exception
    HTTPError = _RequestsHttpError

    def request(self, method, url, **kwargs):
        return _requests_request(method, url, **kwargs)

    def get(self, url, **kwargs):
        return _requests_request("GET", url, **kwargs)

    def post(self, url, **kwargs):
        return _requests_request("POST", url, **kwargs)

requests = _RequestsModule()
if sys is not None:
    try:
        sys.modules["requests"] = requests
    except Exception:
        pass

scope = {"BASE_URL": BASE_URL}
scope["requests"] = requests
scope["http_get"] = http_get
scope["http_post"] = http_post
try:
    exec(code, scope, scope)
    _result = scope.get("result")
    if _result is None:
        print("__MCP_RESULT__")
    else:
        try:
            import json as _json
            print("__MCP_RESULT_JSON__" + _json.dumps(_result, ensure_ascii=False, default=str))
        except Exception:
            print("__MCP_RESULT__" + str(_result))
except Exception as ex:
    print("__MCP_ERROR__" + str(ex))
    raise
""";

        ExecutionResult result = sandbox.Run(wrapper);
        if (!result.Success)
        {
            throw new InvalidOperationException(
                $"Hyperlight Python execution failed with exit code {result.ExitCode}: {result.Stderr}");
        }

        return result;
    }

    private object? ExtractFinalValue(ExecutionResult result)
    {
        string lastLine = result.Stdout
            .Split('\n', StringSplitOptions.RemoveEmptyEntries | StringSplitOptions.TrimEntries)
            .LastOrDefault() ?? string.Empty;

        if (lastLine.StartsWith("__MCP_RESULT__", StringComparison.Ordinal))
        {
            string value = lastLine["__MCP_RESULT__".Length..];
            return string.IsNullOrEmpty(value) ? null : value;
        }

        if (lastLine.StartsWith("__MCP_RESULT_JSON__", StringComparison.Ordinal))
        {
            string value = lastLine["__MCP_RESULT_JSON__".Length..];
            if (string.IsNullOrWhiteSpace(value))
            {
                return null;
            }

            try
            {
                return JsonSerializer.Deserialize<object>(value);
            }
            catch
            {
                return value;
            }
        }

        if (lastLine.StartsWith("__MCP_ERROR__", StringComparison.Ordinal))
        {
            return lastLine;
        }

        return result.Stdout.TrimEnd();
    }

    private bool ContainsForbiddenHardcodedApiUsage(string code)
    {
        if (allowedBaseUris.Count == 0)
        {
            return false;
        }

        MatchCollection matches = AbsoluteHttpUrlRegex.Matches(code);
        foreach (Match match in matches)
        {
            if (!Uri.TryCreate(match.Value, UriKind.Absolute, out Uri? discoveredUrl))
            {
                continue;
            }

            bool isAllowed = allowedBaseUris.Any(baseUri =>
                discoveredUrl.Scheme.Equals(baseUri.Scheme, StringComparison.OrdinalIgnoreCase) &&
                discoveredUrl.Host.Equals(baseUri.Host, StringComparison.OrdinalIgnoreCase) &&
                discoveredUrl.Port == baseUri.Port &&
                discoveredUrl.AbsoluteUri.StartsWith(baseUri.AbsoluteUri, StringComparison.OrdinalIgnoreCase));

            if (!isAllowed)
            {
                return true;
            }
        }

        return false;
    }

    private static IReadOnlyList<Uri> NormalizeAllowedBaseUris(IReadOnlyList<string>? input)
    {
        if (input is null || input.Count == 0)
        {
            return [];
        }

        HashSet<string> seen = new(StringComparer.OrdinalIgnoreCase);
        List<Uri> normalized = [];

        foreach (string value in input)
        {
            if (!Uri.TryCreate(value, UriKind.Absolute, out Uri? uri))
            {
                continue;
            }

            if (uri.Scheme != Uri.UriSchemeHttp && uri.Scheme != Uri.UriSchemeHttps)
            {
                continue;
            }

            Uri withSlash = EnsureTrailingSlash(uri);
            if (!seen.Add(withSlash.AbsoluteUri))
            {
                continue;
            }

            normalized.Add(withSlash);
        }

        return normalized;
    }

    private static Uri EnsureTrailingSlash(Uri uri)
    {
        if (uri.AbsoluteUri.EndsWith("/", StringComparison.Ordinal))
        {
            return uri;
        }

        return new Uri(uri.AbsoluteUri + "/", UriKind.Absolute);
    }

    private void ThrowIfDisposed()
    {
        if (disposed)
        {
            throw new ObjectDisposedException(nameof(HyperlightSandboxRunner));
        }
    }

    public void Dispose()
    {
        if (disposed)
        {
            return;
        }

        disposed = true;
        sandbox.Dispose();
    }

}

Enter fullscreen mode Exit fullscreen mode

And finally, we need to register it with the SandboxRunnerFactory:

public static class SandboxRunnerFactory
{
    public static ISandboxRunner Create(
        IConfiguration configuration,
        ILoggerFactory loggerFactory,
        IReadOnlyList<string>? allowedBaseUrls = null)
    {
        ArgumentNullException.ThrowIfNull(configuration);
        ArgumentNullException.ThrowIfNull(loggerFactory);

        string runnerName = configuration["CodeMode:Runner"]?.Trim() ?? "auto";
        int timeoutMs = configuration.GetValue<int>("CodeMode:TimeoutMs", 5000);
        int maxToolCalls = configuration.GetValue<int>("CodeMode:MaxToolCalls", 10);

        if (string.Equals(runnerName, "opensandbox", StringComparison.OrdinalIgnoreCase))
        {
            string? domain = configuration["OpenSandbox:Domain"] ?? "localhost:8080";

            OpenSandboxRunnerOptions options = new()
            {
                Domain = domain,
                ApiKey = configuration["OpenSandbox:ApiKey"],
                Image = configuration["OpenSandbox:Image"] ?? "python:3.12-slim",
                TimeoutSeconds = configuration.GetValue<int>("OpenSandbox:TimeoutSeconds", 5 * 60),
                ReadyTimeoutSeconds = configuration.GetValue<int>("OpenSandbox:ReadyTimeoutSeconds", 30),
                RequestTimeoutSeconds = configuration.GetValue<int>("OpenSandbox:RequestTimeoutSeconds", 30),
                UseServerProxy = configuration.GetValue<bool>("OpenSandbox:UseServerProxy", true),
                Timeout = TimeSpan.FromMilliseconds(timeoutMs),
                MaxToolCalls = maxToolCalls,
                AllowedBaseUrls = allowedBaseUrls,
            };

            return new OpenSandboxRunner(options, loggerFactory, allowedBaseUrls);
        }

        if (string.Equals(runnerName, "local", StringComparison.OrdinalIgnoreCase))
        {
            return new LocalConstrainedRunner(
                timeout: TimeSpan.FromMilliseconds(timeoutMs),
                maxToolCalls: maxToolCalls,
                loggerFactory.CreateLogger<LocalConstrainedRunner>(),
                allowedBaseUrls);
        }

        try
        {
            return new HyperlightSandboxRunner(
                timeout: TimeSpan.FromMilliseconds(timeoutMs),
                maxToolCalls: maxToolCalls,
                loggerFactory.CreateLogger<HyperlightSandboxRunner>(),
                allowedBaseUrls);
        }
        catch (Exception ex)
        {
            loggerFactory.CreateLogger("McpServer.CodeMode.SandboxRunnerFactory")
                .LogWarning(ex, "Falling back to local sandbox runner because Hyperlight runner could not be created.");

            return new LocalConstrainedRunner(
                timeout: TimeSpan.FromMilliseconds(timeoutMs),
                maxToolCalls: maxToolCalls,
                loggerFactory.CreateLogger<LocalConstrainedRunner>(),
                allowedBaseUrls);
        }
    }

}

Enter fullscreen mode Exit fullscreen mode

Test it all

In the root of the project, run:

aspire run

Enter fullscreen mode Exit fullscreen mode

Then access to https://localhost:17294, and click on mcp-server, you should see:

demo 1

Open up the web, then type use codemode to get random brewery

demo 2

Back to the mcp-server logs:

demo 3

And the web:

demo 4

Try with the complex scenario:

The prompt:

Write Python code for execute mode.

Requirements:
1) Get breweries in San Diego.
2) Return the total number of breweries found for San Diego.
3) Search for breweries matching "moon" and return top 5 with only:
   - name
   - city
4) Also compute a San Diego summary:
   - count by brewery_type
   - top 3 cities by number of breweries (if city is missing, use "Unknown")
5) Deduplicate breweries by id before counting.
6) Sort outputs deterministically:
   - type summary: count desc, then type asc
   - top cities: count desc, then city asc
   - moon top 5: name asc
7) Return one JSON object in `result` with this exact shape:
{
  "san_diego_total": number,
  "san_diego_type_summary": [{"brewery_type": string, "count": number}],
  "san_diego_top_cities": [{"city": string, "count": number}],
  "moon_top_5": [{"name": string, "city": string}]
}
8) If any tool call fails, set a top-level "errors" array with readable messages but still return partial data.

Enter fullscreen mode Exit fullscreen mode

The logs are more complex, see:

info: McpServer.CodeMode.CodeModeHandlers[0]
      [codemode] Execute executing code (2707 chars).
info: McpServer.CodeMode.Hyperlight.HyperlightSandboxRunner[0]
      Hyperlight constrained execution requested. Timeout: 00:01:30, MaxToolCalls: 10.
info: McpServer.CodeMode.Hyperlight.HyperlightSandboxRunner[0]
      Generated Python code for Hyperlight Execute:
import json
from collections import Counter, defaultdict

errors = []


def fetch_json(url):
    resp = http_get(url)
    if resp["status"] >= 400:
        raise Exception(f"HTTP {resp['status']} for {url}")
    return resp["body"]


def as_list(body):
    if isinstance(body, list):
        return body
    if isinstance(body, dict) and "items" in body and isinstance(body["items"], list):
        return body["items"]
    return []


def dedupe_by_id(items):
    seen = set()
    out = []
    for item in items:
        brewery_id = item.get("id")
        key = brewery_id if brewery_id is not None else json.dumps(item, sort_keys=True, default=str)
        if key in seen:
            continue
        seen.add(key)
        out.append(item)
    return out


san_diego_items = []
moon_items = []

try:
    page = 1
    per_page = 200
    while True:
        body = fetch_json(f"{BASE_URL}/breweries?by_city=San_Diego&page={page}&per_page={per_page}")
        items = as_list(body)
        if not items:
            break
        san_diego_items.extend(items)
        if len(items) < per_page:
            break
        page += 1
except Exception as e:
    errors.append(f"San Diego breweries fetch failed: {e}")

try:
    page = 1
    per_page = 200
    while True:
        body = fetch_json(f"{BASE_URL}/breweries/search?query=moon&page={page}&per_page={per_page}")
        items = as_list(body)
        if not items:
            break
        moon_items.extend(items)
        if len(items) < per_page:
            break
        page += 1
except Exception as e:
    errors.append(f"Moon search failed: {e}")

san_diego_items = dedupe_by_id(san_diego_items)
moon_items = dedupe_by_id(moon_items)

sd_type_counts = Counter()
sd_city_counts = Counter()
for item in san_diego_items:
    brewery_type = item.get("brewery_type") or "Unknown"
    city = item.get("city") or "Unknown"
    sd_type_counts[brewery_type] += 1
    sd_city_counts[city] += 1

san_diego_type_summary = [
    {"brewery_type": brewery_type, "count": count}
    for brewery_type, count in sorted(sd_type_counts.items(), key=lambda kv: (-kv[1], kv[0]))
]

san_diego_top_cities = [
    {"city": city, "count": count}
    for city, count in sorted(sd_city_counts.items(), key=lambda kv: (-kv[1], kv[0]))[:3]
]

moon_top_5 = [
    {"name": item.get("name") or "", "city": item.get("city") or ""}
    for item in sorted(moon_items, key=lambda item: ((item.get("name") or ""), (item.get("city") or "")))[:5]
]

result = {
    "san_diego_total": len(san_diego_items),
    "san_diego_type_summary": san_diego_type_summary,
    "san_diego_top_cities": san_diego_top_cities,
    "moon_top_5": moon_top_5,
}

if errors:
    result["errors"] = errors
info: McpServer.CodeMode.CodeModeHandlers[0]
      [codemode] Execute completed successfully. HasResult: True.
info: Microsoft.AspNetCore.Routing.EndpointMiddleware[1]
      Executed endpoint 'MCP Streamable HTTP | HTTP: POST /mcp/'
dbug: Microsoft.AspNetCore.Server.Kestrel.Connections[9]
      Connection id "0HNLELGG9J670" completed keep alive response.
info: Microsoft.AspNetCore.Hosting.Diagnostics[2]
      Request finished HTTP/1.1 POST http://localhost:5100/mcp - 200 - text/event-stream 1759.5000ms

Enter fullscreen mode Exit fullscreen mode

You can see the LLM model write the Python code, based on the prompt above:

import json
from collections import Counter, defaultdict

errors = []


def fetch_json(url):
    resp = http_get(url)
    if resp["status"] >= 400:
        raise Exception(f"HTTP {resp['status']} for {url}")
    return resp["body"]


def as_list(body):
    if isinstance(body, list):
        return body
    if isinstance(body, dict) and "items" in body and isinstance(body["items"], list):
        return body["items"]
    return []


def dedupe_by_id(items):
    seen = set()
    out = []
    for item in items:
        brewery_id = item.get("id")
        key = brewery_id if brewery_id is not None else json.dumps(item, sort_keys=True, default=str)
        if key in seen:
            continue
        seen.add(key)
        out.append(item)
    return out


san_diego_items = []
moon_items = []

try:
    page = 1
    per_page = 200
    while True:
        body = fetch_json(f"{BASE_URL}/breweries?by_city=San_Diego&page={page}&per_page={per_page}")
        items = as_list(body)
        if not items:
            break
        san_diego_items.extend(items)
        if len(items) < per_page:
            break
        page += 1
except Exception as e:
    errors.append(f"San Diego breweries fetch failed: {e}")

try:
    page = 1
    per_page = 200
    while True:
        body = fetch_json(f"{BASE_URL}/breweries/search?query=moon&page={page}&per_page={per_page}")
        items = as_list(body)
        if not items:
            break
        moon_items.extend(items)
        if len(items) < per_page:
            break
        page += 1
except Exception as e:
    errors.append(f"Moon search failed: {e}")

san_diego_items = dedupe_by_id(san_diego_items)
moon_items = dedupe_by_id(moon_items)

sd_type_counts = Counter()
sd_city_counts = Counter()
for item in san_diego_items:
    brewery_type = item.get("brewery_type") or "Unknown"
    city = item.get("city") or "Unknown"
    sd_type_counts[brewery_type] += 1
    sd_city_counts[city] += 1

san_diego_type_summary = [
    {"brewery_type": brewery_type, "count": count}
    for brewery_type, count in sorted(sd_type_counts.items(), key=lambda kv: (-kv[1], kv[0]))
]

san_diego_top_cities = [
    {"city": city, "count": count}
    for city, count in sorted(sd_city_counts.items(), key=lambda kv: (-kv[1], kv[0]))[:3]
]

moon_top_5 = [
    {"name": item.get("name") or "", "city": item.get("city") or ""}
    for item in sorted(moon_items, key=lambda item: ((item.get("name") or ""), (item.get("city") or "")))[:5]
]

result = {
    "san_diego_total": len(san_diego_items),
    "san_diego_type_summary": san_diego_type_summary,
    "san_diego_top_cities": san_diego_top_cities,
    "moon_top_5": moon_top_5,
}

if errors:
    result["errors"] = errors

Enter fullscreen mode Exit fullscreen mode

And when you're back to the web:

demo 6

Conclusion

Hyperlight changes the execution model from infrastructure-driven sandboxing to runtime-embedded isolation. Instead of delegating code execution to external container services over network boundaries, MCP CodeMode can invoke Hyperlight directly inside the host process while still enforcing hypervisor-backed isolation.

This collapses latency, reduces operational complexity, and removes an entire class of distributed failure modes, while preserving strong containment for untrusted agent workloads. It is not a replacement for container platforms like OpenSandbox, but a different execution primitive — optimised for fast, secure, function-level agent execution inside the application runtime itself.

The source code can be found at mcp-experiments

The code runs well on Linux and Windows, but on macOS, the Hyperlight team is still working on support.

Happy hacking!