Agent integration

Connect an MCP client#

The public endpoint is https://rogueon.ai/mcp. Choose Streamable HTTP in your MCP client. It requires no API key, account, OAuth flow, or model subscription to read the documentation. Let the client negotiate the protocol; the endpoint supports the 2026-07-28 protocol and compatible 2025 clients.

For Cursor, merge the following entry into your project's .cursor/mcp.json, preserving any existing servers. Reload the MCP configuration and confirm that list_docs, search_docs, and get_doc are listed. Other editors may use a different configuration format for the same endpoint.

{
  "mcpServers": {
    "rogue-docs": {
      "url": "https://rogueon.ai/mcp"
    }
  }
}

• Cursor MCP configuration reference

Claude Code#

Run this in the project where you want the documentation available. The command registers the remote server; it does not generate or publish an app. Use /mcp in Claude Code to inspect the connection. If rogue-docs is already configured, inspect or update that entry instead of adding a duplicate.

claude mcp add --transport http rogue-docs https://rogueon.ai/mcp

• Claude Code HTTP MCP reference

Tool contract#

All tools accept locale: en or pt-br, defaulting to en. Tool results include source URLs. Markdown resources are also discoverable through resources/list and readable through resources/read.

{
  "slug": "project-architecture",
  "locale": "en",
  "section": "runtime-services"
}

Search → retrieve → verify#

1. 01

   Find the relevant source

   Use search_docs with a focused query. Search matches all words, ignores accents, and returns excerpts; an excerpt is not the whole reference.

2. 02

   Read the smallest useful section

   Use the document slug and section ID with get_doc. Omit section when surrounding context is necessary. Keep conditions about artifact format, runtime and availability in any summary.

3. 03

   Retain source and revision

   Cite the returned canonical URL and retain the SHA-256 with cached context. Check the manifest before reusing old content. The hash describes content identity, not a product release or update date.

Before changing a Rogue-generated project:
1. Identify the actual artifact format from its files.
2. Search Rogue docs for the relevant runtime or export contract.
3. Retrieve that section with get_doc; retain its canonical URL.
4. Distinguish documented behavior from assumptions about this project.
5. Implement only the requested change and validate its acceptance criteria.
6. Report the source, changed files, verification, and unresolved dependencies.

Retrieve Markdown over HTTP#

Every article has an explicit /markdown URL. HTML and Markdown are generated from the same source; the Markdown response declares the canonical HTML URL. This path works with a normal HTTP client when MCP is unavailable or unnecessary.

curl -fsSL https://rogueon.ai/docs/project-architecture/markdown

• Architecture · Markdown

Cache by content hash#

The JSON manifest has schemaVersion: 1 and lists each document's slug, url, markdownUrl, sha256, byte count and section IDs. Document hashes cover the exact UTF-8 Markdown response body; section hashes cover that section's Markdown. The HTTP ETag identifies the manifest representation and supports If-None-Match with a 304 response.

curl -fsS -D - https://rogueon.ai/docs/manifest.json

• Open documentation manifest

Discovery and complete corpus#

llms.txt is the compact discovery index. llms-full.txt contains the public guides in both languages. Prefer a page or section for routine retrieval; use the full corpus when an offline task needs broader coverage. These files are discovery resources, not crawler permission rules or a guarantee of ingestion.

• llms.txt
• llms-full.txt

Transport, limits and scope#

Tool calls are read-only and do not invoke a language model. Your coding agent still has its own provider usage and permissions. An explicit browser Origin must be on the server allowlist; ordinary server-side MCP clients can connect without an Origin header. Recheck configuration and protocol errors before retrying, rather than adding credentials this endpoint does not require.