Vibe Coding
The Vibe Coding platform enables end users to create, modify, and publish applications through natural language. The platform is responsible for understanding requirements, generating code, running verification, and continuously iterating, ultimately delivering applications that can be accessed long-term. Products such as Vercel v0 and Lovable all belong to this category.
EdgeOne Makers integrates core capabilities such as production, deployment, compliance detection, access, and multi-tenant isolation into a single platform, enabling platform providers to build a complete Vibe Coding pipeline from conversational generation to real-time preview, application launch, and multi-tenant isolation without building or maintaining the underlying infrastructure.
Solution Value
Benefits | Specific Manifestation |
Integrated production environment | Agent orchestration, model invocation, session state, and sandbox tool access are closed-loop within the same runtime, reducing cross-service assembly and Ops workload. |
Isolation and lifecycle balance | Workspaces are isolated by session, while project code is retained through external persistence, balancing runtime security and continuous development experience. |
Unified cloud-edge delivery | Unified hosting of static, dynamic, full-stack, and Agent applications, with workloads running on the edge or in the cloud based on application form. |
Automatic project review | Provides project-level content security detection and risk handling capabilities, and notifies platform parties through event mechanisms when risks change. |
Overall Architecture: 1 Vibe Coding Platform + N Project Applications
The production-side Coding Agent generates code by invoking large models and using sandbox tools. The deployment side publishes the generated applications to the Makers cloud-edge integrated runtime environment. The review side performs multi-layer compliance checks on live pages, deployment artifacts, and domains. The access side achieves low-latency access and security protection through edge access. Multi-tenant isolation ensures that each AI-generated project is deployed and runs independently.
Core Capabilities
Production Side: Coding Agent
The Agent runtime handles large model invocation and business decision-making, and drives the Sandbox workspace to perform file read/write operations, dependency installation, and code execution. It generates accessible preview links for users to continue conversations, provide feedback, and iterate. The production process also supports multi-model selection and full-chain observability of conversation storage, ensuring that projects can be sustainably developed, restored, and tracked.
Agent Runtime
It hosts LLM invocation, Agent loop orchestration, and business logic, supporting long-running tasks, session-based routing, and automatic scaling. Each public file in the
agents/ directory corresponds to an Agent endpoint, and the platform handles routing and injects runtime context:// agents/chat.tsexport async function onRequest(context: any) {context.store // Conversation storagecontext.tools // Sandbox tools, from the LLM's perspectivecontext.sandbox // Sandbox atomic APIs, from the developer's perspective}
Feature | Description |
Long-running Task | Supports multi-turn tool invocation, dependency installation, and build processes for Agents |
Conversation Routing | The conversation_id binds the dialog context to the sandbox workspace, enabling multiple rounds of requests to reuse the same project. |
Framework Replaceable | Can integrate with frameworks such as DeepAgents, LangGraph, CrewAI, OpenAI Agents SDK, and Claude Agent SDK. |
Separation of Duties | agents/ hosts AI logic; cloud-functions/ hosts regular business APIs. |
Integrating Models
Makers Models provides a unified access channel from client SDKs to major model providers. Developers can invoke models from multiple mainstream providers through a unified endpoint and a single API Key, without needing to adapt to provider-specific interfaces or manage multiple sets of calling protocols.
Method | Configuration | Scenarios |
Built-in Model | Configure the Makers Models API Key and gateway address, and use the model quota provided by the platform. | No external model provider account required, suitable for rapid validation. |
Bring Your Own API Key | Bind your own model provider. | Requires specific models or prefers autonomous control over costs and quotas. |
Sandbox Tools
Sandbox tools use isolated Tencent Cloud instances under the hood to handle the "side effects" of Agent execution, such as running commands, reading and writing files, controlling browsers, and executing code. This enables LLMs to not only output text but also perform actual development operations.
Capability | Description |
Isolate sessions | Each session uses an independent workspace, and exceptions should not affect other tenants. |
Create on demand | Created only when first needed for execution and automatically recycled after timeout. |
Execute commands and operate files | Install dependencies, build, start services, and read/write project files |
Preview ports | Map services within the sandbox to temporary public network addresses. |
Persist data | Persist generated code via Blob to ensure retention across conversations and instances. |
Control instances | Query status, extend timeout, obtain address, or actively terminate an instance |
Real-Time Preview
Preview is central to the Vibe Coding experience. Users need to see the result after code is generated. The preview address is mapped from a sandbox port and becomes invalid when the sandbox is reclaimed.
Sandbox Persistence
A sandbox is a temporary workspace that is reclaimed upon timeout. To support users in resuming development across different time periods, project files must be persisted. Therefore, when you need to continue within the same session across different sandbox instances: call
persist() to save a snapshot, and after the instance is rebuilt, call restore() to restore it.
Save the workspace code.
Call
persist() after each round of code modifications is completed:try {const persist = await context.sandbox.persist();} catch (error) {// A save failure does not affect the current sandbox's operation. You can log the failure and retry later.console.warn('Failed to save workspace:', error);}
persist() does not require passing in conversation_id. The current sandbox is already bound to the conversation, and each call updates the latest snapshot of that conversation. It is recommended to call it at the following times:After each round of code modifications is completed;
Save periodically during long-running task execution.
Restore the workspace code.
Call
restore() after a new sandbox is created and before the Agent starts reading and writing project files:const result = await context.sandbox.restore();
Note: If
restore() returns failed, do not call persist() again in this round. Otherwise, you may overwrite the last usable snapshot with an incomplete workspace.Deployment Side: Cloud-Edge Integration
For Vibe Coding platforms, Makers provides a scalable, deliverable, and cloud-edge integrated production runtime environment for AI-generated multi-form applications.
Application Form | Typical Technology | Runtime and Delivery Mode |
Frontend application | React,Vue,SSG,SPA | Static resource hosting with caching and acceleration through edge networks |
Full-stack framework | Next.js,Nuxt,SSR,ISR,PPR | Static content is delivered through static resource hosting, and dynamic logic is handled by cloud functions. |
Backend capabilities | Functions, APIs, and scheduled tasks | Running on Cloud Runtime with elastic scaling per request |
Agent application | Claude,OpenAI,LangGraph,CrewAI | Running on Agent Runtime and using platform capabilities such as models, sandboxes, and storage |
Automatic Project Review
Makers provides Vibe Coding platforms with project-level content security detection and risk handling capabilities, completing review before works are published and integrating with platforms through an event mechanism when risks change.
Makers is responsible for platform-side security detection and risk handling, while the platform side is responsible for work presentation, business operations, and mapping review statuses to its own business statuses.
Collaboration Capability | Function |
Unified SDK Access | Integrate deployment, status query, and disposal requests into the platform's existing release process. |
Standard Webhook Collaboration | Proactively sync review results and risk changes to avoid continuous polling by the platform. |
Disposal Within Authorized Scope | Perform review, restoration, blocking, and deletion on a per-project basis to reduce the risk of misoperations. |
Access Side: Edge Access and Security
After the project goes live, requests are routed to the nearest EdgeOne global edge nodes (3,200+) for access, enabling both access acceleration and security protection to take effect simultaneously.
Multi-Tenant Isolation: Multiple Projects per Tenant
Makers adopts a model where one platform tenant manages multiple projects, with each AI project independently deployed and running. For details, see Platform and Multi-Tenancy.
Open Source Project Examples
vibe-coding-agent is a code generation template implemented by Makers based on the Claude Agent SDK + TypeScript. You can use it to quickly build your own Vibe Coding platform.
Out-of-the-Box Capabilities
Capability | Current Implementation |
Agent Runtime | Task orchestration and tool invocation using Claude Agent SDK |
Model Invocation | Switch between multiple large models for invocation based on Makers Models. |
Code Sandbox and Real-Time Preview | Code generated by LLMs runs securely in an isolated environment. The sandbox can obtain a preview URL, allowing users to see the execution results immediately after code generation. |
Persist workspace | Restore the workspace from a source code snapshot, reinstall dependencies, and restart the preview. |
Browse files | The /file route reads workspace text files on demand, and the frontend displays the file tree and source code. |
Export source code | Package the workspace and download the source code archive. After GitHub OAuth is configured, it can be pushed to the user's repository. |
Getting Started
npm install -g edgeoneedgeone makers create --template vibe-coding-agentcd vibe-coding-agent && edgeone makers dev# runtime: http://localhost:8088# devtools: http://localhost:8088/agent-metrics
The deployment capability is currently in a grayscale testing phase. To integrate this capability into your product, contact us at Contact Us to request access. We will provide open APIs and technical support, and assist with solution implementation based on your specific business scenarios.
