For AI agents: The documentation index is at https://docs.digitalocean.com/llms.txt. Markdown versions of pages use the same URL with index.html.md in place of the HTML page (for example, append index.html.md to the directory path instead of opening the HTML document).

The DigitalOcean Documentation MCP server gives AI coding assistants read-only access to the public DigitalOcean documentation. Assistants can search the docs, read pages as Markdown, and find quickstarts and troubleshooting articles while you work. The server is hosted at https://docs.mcp.digitalocean.com/mcp, uses the Streamable HTTP transport, and does not require a DigitalOcean account or API token.

To manage DigitalOcean resources as well as read the docs, connect the MCP Router instead. The router includes the documentation tools.

Connect the Documentation MCP Server

Add the server to your MCP client with the following URL. The server does not use authentication, so do not add an Authorization header.

https://docs.mcp.digitalocean.com/mcp

Claude Code

Run the following command to add the server to Claude Code for all of your projects:

claude mcp add --transport http --scope user digitalocean-docs https://docs.mcp.digitalocean.com/mcp

To add the server to the current project only, omit --scope user. Run claude mcp list and confirm that digitalocean-docs shows a Connected status. For more options, see Connect Claude Code to tools via MCP.

Claude and Claude Desktop

Add https://docs.mcp.digitalocean.com/mcp as a custom connector. For steps, see Get started with custom connectors using remote MCP.

Cursor

Use the following button to add the server to Cursor:

You can also add the server manually. Open ~/.cursor/mcp.json to make the server available in every project, or .cursor/mcp.json to add it to one project, and add the following entry under mcpServers:

{
  "mcpServers": {
    "digitalocean-docs": {
      "url": "https://docs.mcp.digitalocean.com/mcp"
    }
  }
}

For more information, see Cursor’s MCP documentation.

Visual Studio Code

Open the Command Palette and run MCP: Add Server. Choose HTTP, enter https://docs.mcp.digitalocean.com/mcp, and enter digitalocean-docs as the server ID. To use the server across workspaces, choose Global, then Copilot Global when prompted. For other configuration destinations, see Add MCP servers in VS Code.

Codex

Run the following command to add the server to Codex:

codex mcp add digitalocean-docs --url https://docs.mcp.digitalocean.com/mcp

To add the server manually instead, open ~/.codex/config.toml and add the following table:

[mcp_servers.digitalocean-docs]
url = "https://docs.mcp.digitalocean.com/mcp"

Run codex mcp list to confirm that Codex lists the server. For more options, see the Codex MCP documentation.

Other Clients

Other MCP clients that support remote servers can connect with the same URL. Add https://docs.mcp.digitalocean.com/mcp as a Streamable HTTP server using the method in your client’s documentation.

Verify the Connection

Ask your assistant a question that names the server, such as the following prompt:

Use the DigitalOcean Documentation MCP server to find the App Platform quickstart and summarize its steps. Cite the page URL.

The assistant calls tools such as docs-search or docs-get-quickstart and cites docs.digitalocean.com URLs in its answer. If it answers without calling a tool, check that your client lists the server and that its tools are enabled.

This content is automatically generated from https://github.com/digitalocean-labs/mcp-digitalocean/blob/main/pkg/registry/docs/README.md.

Documentation Tools

Read-only tools for querying DigitalOcean’s public documentation. No authentication required.

Supported Tools

• docs-search Keyword search over the titles, descriptions, and URLs of every public docs page (~4,600 pages, including individual support articles and API/CLI/SDK reference pages). Returns ranked results with canonical URL, product, page type, last-updated date, and an excerpt: the passage of the page body that mentions the query terms when available, otherwise the page description. Ranking does not search page bodies. Arguments:

  • Query (string, required): Search keywords
  • Limit (number, default: 10): Maximum number of results to return (1 to 50)
  • Product (string, optional): Restrict results to one product or area (e.g., "droplets", "app platform", "vpc", "support", "api")

• docs-get-page Fetch the markdown content of a specific docs page, whole or in windows. Arguments:

  • URL (string, required): Full URL or path of the docs page (e.g., https://docs.digitalocean.com/products/droplets/getting-started/quickstart/ or /products/droplets/getting-started/quickstart/)
  • MaxLength (number, optional): Maximum UTF-8 bytes to return (integer from 0 to 10,485,760); omit or use 0 for the whole page. Windows end at character boundaries. A budget too small for the next character returns an argument error. The truncation note gives the StartIndex to continue from.
  • StartIndex (number, optional): UTF-8 byte offset to start from (integer from 0 to 10,485,760, default 0). An offset inside a character advances to the next character

• docs-find-for-service List documentation pages for a given DigitalOcean service. Arguments:

  • Service (string, required): DigitalOcean service name (e.g., "droplets", "kubernetes", "app platform", "databases")

• docs-get-quickstart Get the quickstart or getting-started guide for a service. Arguments:

  • Service (string, required): DigitalOcean service name (e.g., "droplets", "kubernetes", "app platform")

• docs-troubleshoot Find support and troubleshooting articles matching an error message or symptom by keyword, ranking support articles and troubleshooting-style titles first. Returns the content of the best matching page (up to 16 KB, with a pointer to the rest), plus related pages with excerpts. Arguments:

  • Symptom (string, required): Error message, symptom description, or keywords (e.g., "520 status code", "deployment failed health check")
  • Limit (number, default: 5): Maximum number of results to return (1 to 50)

• docs-get-related Extract and categorize all outbound documentation links from a specific docs page. Returns links grouped by type (how-to, reference, support, getting-started, concept, details). Arguments:

  • URL (string, required): Full URL or path of the docs page to extract links from

How It Works

• Requests, including redirects, are confined to https://docs.digitalocean.com; page URLs are canonicalized; the cache is capped at 256 entries; concurrent fetches are coalesced; index text is flattened; off-origin related links are dropped; tool descriptions tell agents not to follow instructions in page text.

• docs-search and docs-troubleshoot rank pages from docs.digitalocean.com/llms-index.json, the published per-page metadata index (title, description, product, section, canonical URL, last-modified). It covers every public page, including support articles that the curated llms.txt overview omits. If a JSON refresh fails, the client keeps a previously loaded JSON index and retries after 10 minutes. With no previous JSON index it falls back to llms.txt; a failed refresh of that fallback also keeps the last good index.

• Ranking is keyword-based: query terms are lowercased, stripped of punctuation and stop words, and matched against the tokens of the title, description, product, and URL path: a term matches a token exactly, by light stem (configuration/configure), or as a prefix of at least four characters (postgres/postgresql), never as an arbitrary substring (spaces does not match namespace). Fields are tokenized once when the index loads, so a query costs about 10 ms over 4,600 pages. Title matches receive more weight than description and URL matches, pages matching every term receive an additive boost, and terms that match very few pages (such as an error code) weigh double so they are not drowned out by common product words. Page bodies are not used for ranking; see corpus-search-followup.md for what that would take.

• Excerpts come from docs-corpus.jsonl, the published page-body feed. After ranking, the client schedules only the shards covering the returned pages, in result order (at most 8 per query, at most 4 corpus loads at a time across the client, 5 s enrichment deadline, 30 min shard cache). Concurrent queries share manifest and shard loads. A caller can stop waiting without cancelling another query’s load; shared loads have their own 5 s deadline, and cancellations are not negative-cached. Each excerpt preserves source case and UTF-8 boundaries, considers the first 8 occurrences of each query term within the first 64 KB of the body, and chooses a window of up to 300 bytes plus ellipses. Whitespace snapping preserves a complete query hit. Missing, oversized, slow, or unmatched content falls back to the description; the corpus is never loaded in full.

• Product filtering prefers exact product names, service aliases, and URL slugs. If none match, it uses whole section tokens with light stemming and prefixes of at least four characters, so short strings do not match arbitrary substrings.

• docs-find-for-service uses per-service llms.txt files

• Uses markdown_url from the cached page index when available, including the last good index during refresh failures. Falls back to the index.html.md endpoint and then the original page URL without loading the index for a page read; only docs.digitalocean.com URLs are fetched. All requests and redirects must stay on the HTTPS docs.digitalocean.com origin. Each HTTP request has a 15 s timeout; oversized bodies are rejected even without a reliable Content-Length (10 MB accepted-body cap). Page windows do not bypass this response cap.

• In-memory caching (30 min for pages, 1 hour for indexes); concurrent cold loads of the index share one fetch

• Supports common service name aliases (e.g., “k8s” → “kubernetes”, “gpu” → “bare-metal-gpus”)

• When the docs service is enabled, the server’s initialize response carries instructions noting that prior knowledge of DigitalOcean products may be out of date and pointing to these tools. MCP clients load server instructions into the model’s context at connect time.

Ranking regression set

testdata/regression_queries.json is a judged query set (28 queries: support, how-to, quickstart, API/CLI/SDK reference, and prefix/abbreviation queries such as mongo backups, with expected canonical URLs). TestRegressionSet_Fixture runs it against the 218-page sample in testdata/llms-index.json (each target plus its strongest competitors) and requires every expected page in the top 5; MRR is gated loosely at 0.80 so tie-break changes do not fail the build (measured 0.88). Additional tests reject unrelated substring matches and exercise rarity weighting at 4,600-page scale. The sample does not reproduce the full live universe.

TestLiveRegressionSet (build tag integration, not run in standard CI) runs the set against the live index and requires Recall@5 >= 0.9, MRR >= 0.75, and under 50 ms per query of ranking time; measured 28/28, 0.875, 11 ms. TestLiveContentExcerpts reports excerpt coverage and shard-fetch latency with a fresh corpus cache per query; measured 136/140 results and 28/28 top-1 results with a body excerpt, cold avg 267 ms / max 535 ms, warm 3 ms. Add a judged row before changing ranking.

HTTP search endpoint (preview)

For clients that do not speak MCP, a remote deployment started with --allow-anonymous --services docs --docs-search-api also serves GET /v1/search. It uses the same client, index, ranker, and excerpt layer as docs-search, so the two return the same pages and share caches and the corpus load budget. The server refuses to start with --docs-search-api unless anonymous access is enabled, which already requires the docs-only service set.

Unknown or repeated parameters return 400, so each distinct URL is a distinct search and cannot be used to bypass a CDN cache. Only GET (and CORS preflight OPTIONS) is accepted.

curl 'https://<host>/v1/search?q=deploy+from+container+image&limit=3'

The response is {"results": [...], "meta": {...}}. Each result has title, url, markdown_url, product, page_type, last_modified, and excerpt; meta carries query, product, total, pages_searched, and stability (preview). Errors use the public API shape {"id", "message"} (bad_request, method_not_allowed, too_many_requests, service_unavailable) with Cache-Control: no-store. Successful responses send Access-Control-Allow-Origin: * (no credentials are involved) and Cache-Control: public, max-age=300. Query text is not logged; the request log records term count, result count, zero-result flag, and duration.

Rate limiting is a per-client token bucket (IPv6 grouped by /64), sharded so memory and eviction cost stay bounded:

• Without a trusted header the limiter keys on the connection address, which behind an edge is the edge itself, so enforce refuses to start without one and shadow logs a warning.
• With a trusted header, requests that lack a usable value did not come through the edge; they share a single bucket rather than going unlimited.
• RateLimit-Policy is sent on every response while the limiter is on, because it is the same for all clients. Per-client values (RateLimit-Remaining, RateLimit-Reset, Retry-After) are sent only on 429, which is never cached, so a CDN cannot serve one client’s budget to another.
• The limiter is per instance, so the effective ceiling scales with instance count; it is abuse protection, not a quota. Bulk consumers should read docs-corpus.jsonl instead of paging through search, and the 429 body says so.
• A client is logged once when it first goes over its limit, with its IP or /64, so abusive sources can be blocked at the edge.

Example Usage

• Search documentation: Tool: docs-search Arguments:

  • Query: "kubernetes networking"
  • Limit: 5

• Fetch a specific page: Tool: docs-get-page Arguments:

  • URL: "/products/droplets/getting-started/quickstart/"

• Browse a service: Tool: docs-find-for-service Arguments:

  • Service: "app platform"

• Get a quickstart guide: Tool: docs-get-quickstart Arguments:

  • Service: "databases"

• Diagnose an error: Tool: docs-troubleshoot Arguments:

  • Symptom: "520 status code from my app"
  • Limit: 3

• Find related pages: Tool: docs-get-related Arguments:

  • URL: "/products/app-platform/how-to/manage-domains/"

Notes

• All tools are read-only and do not require a DigitalOcean API token.
• All tools use argument-based input; do not use resource URIs.
• All responses are returned as markdown text.
• Service name aliases are supported (e.g., “k8s” for “kubernetes”, “postgres” for “postgresql”).