MCP
Connect your preferred MCP-compatible development tool to the s&box docs server.
Use MCP when your IDE or agent supports MCP servers.
1) Configure your preferred tool
Choose one of the following setup options:
- Click the install button above.
- Approve the VS Code prompt to add the server.
- Reload VS Code if the MCP tools do not appear immediately.
Choose one of the following setup options:
- Click the install button above.
- Approve the Cursor prompt to add the server.
- Restart Cursor if the MCP tools do not appear immediately.
Choose one of the following setup options:
This installs sdocs as a Claude Desktop extension.
It requires Node.js 18+ on your machine. The .mcpb does not run the
hosted MCP server directly inside Claude Desktop. Instead, it installs a
small local Node.js proxy that forwards Claude Desktop MCP traffic to
the hosted sdocs server over HTTP.
Use this when you want the packaged Claude Desktop install flow. If you want Claude Desktop to connect directly to the hosted MCP endpoint without the local Node.js proxy, use the Custom Connector option below instead.
Download .mcpb- Download the
.mcpbbundle. - Make sure
Node.js 18+is installed locally. - Open the bundle with Claude Desktop.
- Approve the install prompt.
- Restart Claude Desktop if the tools do not appear immediately.
Choose one of the following setup options:
Use this when your IDE or agent supports adding a remote MCP server by URL.
https://sdocs.suiram.dev/api/v1/mcp- Open your tool's MCP or connectors settings.
- Add a new MCP server or connector.
- Choose the remote HTTP or streamable HTTP transport when prompted.
- Paste the
sdocsMCP URL above. - Save the configuration and restart the tool if required.
2) Restart your tool
After saving the config, restart your IDE/agent so it can discover the sdocs MCP server.
3) Verify with a quick prompt
Search for `Sandbox.Component`, inspect its members, then show the exact signature for `OnUpdate`.Advanced
The MCP server now behaves like a typical SDK documentation server rather than a generic docs search endpoint. It exposes a small set of structured tools that let an AI agent progressively discover the s&box API using the standard workflow:
- Search
- Inspect
- Read details
In practice, an agent will usually:
- Call
search_docsto find relevant guides and API entities - Call
read_docon the best returned handle - Inspect the returned references and tips
- Call
read_docagain on any reference that looks relevant - Repeat until it has enough exact API and guide context to answer
This makes the s&box documentation MCP feel closer to a .NET API browser, where namespaces, types, and members are returned in a predictable structure instead of as raw page text.
Transport and Payload Format
The remote MCP endpoint uses standard MCP JSON-RPC over Streamable HTTP. That means installation snippets, request envelopes, tool arguments, and discovery metadata still use JSON where MCP clients expect it.
Tool result text fields and docs:// resource bodies are encoded as compact
TOON to reduce model-facing token usage while preserving the same structured
data model. Resource contents are advertised with mimeType: text/toon.
Compact MCP payloads remove empty fields, avoid repeated raw metadata, and use
short readable keys such as sig, uri, params, returns, and guides.
Tools default to compact responses. Pass detail: "full" in a tool request only
when you need the larger raw field set.
For example, search_docs returns TOON text like this:
results[2]:
- handle: docs://type/Sandbox.Component
kind: class
next: "Call read_doc with this handle to read the API entity and inspect related references."
source: api
summary: "A GameObject can have many components, which are the building blocks of the game."
title: Sandbox.Component
- handle: docs://guide/scene/components
kind: guide
next: "Call read_doc with this handle to read the guide and inspect referenced API symbols."
source: guide
title: Components
query: component update loop
returned: 2API Docs and Guides
The MCP server exposes two connected documentation layers:
- Generated API docs for namespaces, types, and members
- Official Guides fetched live from the s&box docs repository
The relationship is built at runtime. The server scans the upstream guide markdown, looks for symbol mentions in headings, prose, inline code, and code blocks, and then scores which guide pages are most relevant to each API symbol.
That gives agents two useful directions:
- From an API symbol like
Sandbox.Component.OnUpdate(), they can follow related Guides for conceptual context, editor workflows, and practical usage patterns. - From a Guide page, they can inspect the related API symbols that the guide discusses and then open the canonical type or member pages.
Use the API docs for exact contracts. Use Guides for surrounding context.
Tool Reference
Tool responses are returned as TOON text to reduce token usage while preserving the same structured data model. The examples below show tool arguments as JSON because MCP clients send tool arguments through JSON-RPC.
Every tool accepts an optional detail field:
{
"detail": "full"
}Omit it for compact responses. Use detail: "full" only when the compact result
does not include a field you need.
search_docs
Use this first. It searches official Guides and generated API symbols together, including classes, structs, interfaces, enums, constructors, methods, and properties.
Typical input:
{
"query": "component update loop",
"limit": 5
}Typical TOON output includes:
- Mixed guide and API results
- A stable
handlefor each result - The result
kind, title, URL, summary, and source - A
nexttip telling the agent to callread_docon the handle
Example use:
- Find guides and methods related to
Sandbox.Component - Search for
NetworkMode,Ray,SceneTrace, orGameObjectwithout first knowing whether the answer is in a guide, enum, class, method, or property
read_doc
Use this on any handle returned by search_docs or by a previous read_doc
response. The handle may point to a guide, enum, class, method, constructor, or
property.
Typical input:
{
"target": "docs://type/Sandbox.Component"
}Typical TOON output includes:
- The requested guide or API entity
- Exact API details when the target is a symbol
- Guide markdown excerpts when the target is a guide
- Related references with handles
- A
tipon each reference explaining when to read it next - Workflow guidance telling the agent to keep calling
read_docon relevant reference handles
Example use:
- Read
docs://guide/scene/components, then read the referencedSandbox.ComponentandSandbox.Component.OnUpdate()handles - Read
docs://type/Sandbox.NetworkMode, then read related guide handles for networking behavior
Resources
The MCP server also exposes canonical docs:// resources for agents that
already know the exact target they want to load. These resources complement the
tools:
- Use tools for discovery
- Use resources for deterministic page loading
Available resource patterns:
docs://schemadocs://namespace/rootdocs://namespace/{name}docs://type/{full_name}docs://member/{full_name}docs://guide/{path}
Examples:
docs://namespace/Sandbox
docs://type/Sandbox.Component
docs://member/Sandbox.Component.OnUpdate
docs://guide/scene/componentsResources return compact TOON documentation pages. Resource wrappers omit transport-only metadata such as the resource MIME type and repeated URI. Depending on the resource, the payload can include:
- Summary and description
- Exact signatures and declarations
- Parameters, return types, and return documentation
- Examples
- Child namespaces, member listings, related guide links, and related API symbol links
Typical usage:
- Use
search_docsto identify relevant handles - Use
read_docto read the matchingdocs://type/...,docs://member/..., ordocs://guide/...handle - Use the TOON resource payload as the canonical page for that API element or guide
How Agents Use It
For a question like "How do I update a custom component every frame?", a well behaved agent will usually:
- Use
search_docswith a query likecomponent update every frame - Use
read_docon the best returned handle - Inspect the references and their
tipfields - Use
read_docagain on relevant guide, type, or member handles - Repeat until it has enough exact API contracts and guide context to answer
That is the intended MCP interaction model: search broadly, inspect the right symbol, then combine exact API pages with related Guides when the answer needs more than just signatures.