Reference

AI & agents

Building with Dendri through a coding agent? Every docs page ships as plain markdown, the whole site is indexed in llms.txt, and the hosted MCP endpoint at https://dendri.dev/mcp puts the docs one tool call away in Claude Code, opencode, Cursor, or VS Code.

llms.txt

Dendri publishes the llms.txt convention at the site root:

URL	Contents
`/llms.txt`	Curated index: what Dendri is, package names, server routes, and links to every docs page.
`/llms-full.txt`	Every docs page inlined as one markdown file, paste-able into any context window.

If your tool can fetch URLs, pointing it at /llms-full.txt is the fastest way to give it complete, current knowledge of the SDK and server setup.

Markdown mirror of every page

Each docs page is also served as clean markdown at index.md next to the HTML; same content, no chrome, ideal for agents:

curl https://dendri.dev/docs/quickstart/index.md
curl https://dendri.dev/docs/configuration/index.md
curl https://dendri.dev/docs/frameworks/index.md
curl https://dendri.dev/docs/examples/index.md
curl https://dendri.dev/docs/collaboration/index.md
curl https://dendri.dev/docs/api/index.md
curl https://dendri.dev/docs/self-hosting/index.md

The MCP server

Dendri runs a hosted Model Context Protocol endpoint at https://dendri.dev/mcp (Streamable HTTP, stateless, no API key; it serves only these public docs). Point your agent at the URL and it gets the tools below.

Tool	What it returns
`dendri_list_docs`	Every docs page with slug, title, and summary.
`dendri_get_doc`	One full page as markdown (`quickstart`, `api`, `self-hosting`, …).
`dendri_search_docs`	Sections matching a keyword query, ranked, with source URLs.
`dendri_get_example`	A complete, documented example: chat, live cursors, presence, whiteboard, file transfer, video chat, multiplayer game, collaborative editing, or the React/Vue/Svelte integrations.
`dendri_api_reference`	One API area: `dendri`, `room`, `store`, `connections`, `adapters`, `server-api`, `yjs-provider`, `errors`, `entry-points`.
`dendri_get_config_option`	Matching rows of the client option reference for a name or substring (`apiKey`, `turn`, `relay`, …).

Every docs page is also exposed as an MCP resource (dendri://docs/<slug>) for clients that attach resources instead of calling tools.

Claude Code

claude mcp add --transport http dendri-docs https://dendri.dev/mcp

opencode

{
	"$schema": "https://opencode.ai/config.json",
	"mcp": {
		"dendri-docs": {
			"type": "remote",
			"url": "https://dendri.dev/mcp",
			"enabled": true
		}
	}
}

Cursor

{
	"mcpServers": {
		"dendri-docs": {
			"url": "https://dendri.dev/mcp"
		}
	}
}

VS Code

{
	"servers": {
		"dendri-docs": {
			"type": "http",
			"url": "https://dendri.dev/mcp"
		}
	}
}

What agents see, and what they never see

The MCP server and the markdown mirror are generated from these public docs pages and nothing else. They contain:

SDK usage: installing @afterrealism/dendri-client and @afterrealism/dendri-y, every client option, the framework adapters.
Working example code: the same walkthroughs as Building the examples, written and verified against the real SDK.
Deployment how-to: running the server container, flags, TLS, TURN, Redis.

They deliberately do not contain:

Signaling-server source code. The docs describe how to run the published server image, not its internals, so an agent can never paste AGPL server code into your Apache-2.0 (or proprietary) app by accident.
Credentials or endpoints. Dendri is self-host-first: every generated snippet takes your server URL. There is no default host an agent could silently connect you to.

Prompting tips

The examples in these docs are the gold standard: they use the current API and handle cleanup, reconnection, and host migration correctly. Steer your agent toward them:

# With dendri-docs MCP connected:
"Use dendri_get_example('live-cursors') as the reference pattern and build
 a cursor overlay for my React app. My signaling server runs at
 wss://signal.example.com — do not invent any other endpoint."

Pin the pattern, not the vibe. Asking the agent to fetch the relevant doc or example first (one tool call) produces dramatically better Dendri code than letting it improvise from training data.

Agent integration ideas? Open an issue on GitHub.