This MCP server provides AI agents with version-pinned access to NVIDIA CUDA-Q documentation, API references, runnable examples, and execution target information, preventing version skew.
Search documentation: search_docs returns ranked, excerpted documentation results with breadcrumbs and canonical URLs.
Fetch full pages: get_page retrieves a complete documentation page as clean markdown.
Resolve API symbols: find_api pinpoints exact Python or C++ API symbols with kind, doc URL, and excerpt.
Find runnable examples: search_examples locates complete CUDA-Q example programs from the matching release.
List execution targets: list_targets enumerates all 24 CUDA-Q execution targets (simulators, hardware, clouds) with selection guidance.
Version-pinned answers: Defaults to your installed CUDA-Q version and falls back to latest if a specific version is unavailable.
Offline and local: Uses SQLite full-text search with BM25 ranking, no API keys, no embeddings, and works offline once the index is downloaded.
Client integration: Easily integrates with MCP-capable AI clients such as Claude Code, Cursor, and VS Code.
Provides access to NVIDIA CUDA-Q documentation, API reference, and runnable examples, enabling AI agents to search and retrieve quantum computing SDK information with version-pinned accuracy.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@cudaq-docs-mcp Show me a runnable CUDA-Q example for quantum teleportation"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
An MCP server that serves NVIDIA CUDA-Q documentation, API reference, and runnable examples to AI agents: version-pinned to the cudaq you actually have installed.
Community project, not affiliated with or endorsed by NVIDIA. CUDA-Q is a trademark of NVIDIA Corporation.
Quantum SDKs move faster than model training data. Ask an AI assistant to write CUDA-Q code and it answers from whatever it memorized: renamed APIs, retired target names, install steps for a version you do not run. The failure is version skew, and it lands where onboarding matters most: the first ten minutes.
This server gives any MCP-capable agent the current answer instead. Documentation search, exact API symbol resolution, complete runnable examples, and a backend-selection guide, all served from an index of the docs that match your installed cudaq package. No API keys and no embeddings: SQLite full-text search with BM25 ranking, on your machine, offline once the index exists.
Related MCP server: ws-mcp
Register the server with your client; on first use it downloads a prebuilt index (a couple of megabytes) automatically. Building locally is only needed for versions without a prebuilt asset:
uvx cudaq-docs-mcp build --version 0.14.0Claude Code
claude mcp add cudaq-docs -- uvx cudaq-docs-mcpClaude Desktop (claude_desktop_config.json), Cursor (.cursor/mcp.json), or any client that takes a JSON server map:
{
"mcpServers": {
"cudaq-docs": {
"command": "uvx",
"args": ["cudaq-docs-mcp"]
}
}
}VS Code (.vscode/mcp.json):
{
"servers": {
"cudaq-docs": {
"type": "stdio",
"command": "uvx",
"args": ["cudaq-docs-mcp"]
}
}
}Prefer pip? pip install cudaq-docs-mcp and use cudaq-docs-mcp as the command.
Tool | What it returns |
| Ranked doc excerpts with breadcrumbs and canonical URLs |
| One full documentation page as clean markdown |
| Exact Python or C++ symbol, kind, doc URL, and an excerpt |
| Complete runnable programs from the CUDA-Q repository at the matching release |
| All 24 execution targets: simulators, hardware providers, and clouds, with selection snippets and when-to-use guidance |
Resources: cudaq://versions (installed and indexed versions) and cudaq://llms.txt (CUDA-Q's own llms.txt for the served version).
Every tool resolves its docs version in this order:
An explicit version argument ("0.15.0", "latest")
The installed cudaq package, detected from distribution metadata (cudaq is never imported)
latest
Indexes are per-version. When a pinned index is missing the server says so in the response and serves latest instead, with the one command that fixes it. Skew becomes visible instead of silent.
CUDA-Q publishes the raw material: a Sphinx inventory (objects.inv) listing every page and API symbol, markdown mirrors of each docs page, a per-version llms.txt, and example sources in the repository. This server builds on that groundwork:
objects.inv is the crawl manifest and the API symbol table: no scraping heuristics
each markdown mirror is cleaned of theme chrome, code blocks are rebuilt with their language, and heading anchors are preserved for deep links
pages are chunked by heading and indexed in SQLite FTS5 (porter stemming, BM25 ranking)
examples, snippets, and application sources are fetched from the GitHub release tag that matches the docs version
The whole index is one SQLite file per version in your cache directory (cudaq-docs-mcp info shows where). A nightly workflow rebuilds the latest index so refreshes stay a download, not a build.
cudaq-docs-mcp # serve MCP on stdio (what clients run)
cudaq-docs-mcp build # build the index for your installed cudaq, else latest
cudaq-docs-mcp build --version 0.15.0
cudaq-docs-mcp info # cache location, indexed versions, detected cudaqSet CUDAQ_DOCS_MCP_AUTOBUILD=1 to build automatically on first use, and CUDAQ_DOCS_MCP_CACHE to relocate the cache.
Prebuilt indexes for pinned release versions, not just latest
An eval set of real developer questions, with published retrieval scores
CUDA-QX library docs
Issues and PRs are welcome. Commits need a DCO sign-off (git commit -s); see CONTRIBUTING.md. Built in the open with Claude Code.
Apache-2.0. Documentation content belongs to NVIDIA Corporation & Affiliates, originates from the Apache-2.0 licensed NVIDIA/cuda-quantum repository, and every served result links back to the canonical page. See NOTICE.
Resolve a CUDA-Q API symbol to its canonical definition and doc URL.
Args: name: Symbol name, full or partial: "sample", "cudaq.observe", "qvector", "set_target". language: "python" or "cpp" to filter; omit for both. version: Docs version. Defaults to the installed cudaq version.
Inputs: name (required), version, language.
Fetch one documentation page as clean markdown.
Args: path: Page path as returned by search_docs, for example "using/quick_start" or "using/backends/sims/svsims". version: Docs version. Defaults to the installed cudaq version.
Inputs: path (required), version.
List CUDA-Q execution targets (backends) with guidance on choosing.
Covers simulators (CPU, GPU state vector, tensor network, noisy, dynamics), quantum hardware providers, and cloud aggregators, each with selection snippets and doc URLs. Call this when deciding where to run a kernel or when the user names a provider.
Args: category: Optional filter: "simulator", "hardware", or "cloud".
Inputs: category.
Search the NVIDIA CUDA-Q documentation and return ranked excerpts.
Use this before answering any CUDA-Q question from memory: the platform moves quickly and memorized APIs are often stale. Each result carries a breadcrumb, an excerpt, and the canonical doc URL to cite.
Args: query: Natural language or keywords, for example "run kernel on GPU state vector" or "quantinuum credentials". version: Docs version such as "0.15.0" or "latest". Defaults to the installed cudaq version. limit: Maximum number of results (default 5).
Inputs: query (required), limit, version.
Find complete, runnable CUDA-Q example programs.
Sources are the example, snippet, and application files shipped in the CUDA-Q repository at the matching release. Prefer adapting these over writing kernels from memory.
Args: query: What the example should show, for example "GHZ state", "VQE", or "noise model". language: "python" or "cpp" to filter; omit for both. version: Docs version. Defaults to the installed cudaq version. limit: Maximum number of examples (default 3).
Inputs: query (required), limit, version, language.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 5 tools
Each tool targets a distinct resource: docs search, page retrieval, API symbol resolution, example search, and target listing. The two search tools are clearly separated by corpus (docs vs examples), and get_page complements search_docs by fetching full pages.
All tool names follow a consistent verb_noun snake_case pattern: search_docs, get_page, find_api, search_examples, list_targets. The verbs and nouns are descriptive and predictable.
Five tools is well-scoped for a documentation server, covering discovery (search), retrieval (get_page), reference (find_api), examples, and backend selection without unnecessary bloat.
The tool set covers the primary ways users interact with CUDA-Q docs: searching for answers, reading full pages, looking up APIs, finding runnable examples, and selecting execution targets. There are no obvious gaps that would cause dead ends.
Search NVIDIA CUDA documentation and code samples from AI coding agents.
Team docs served to AI agents over MCP - search, Markdown reads, version pinning, read audit.
Agentic search over your Dewey document collections from any MCP-compatible client.
Query any docs site via MCP. Submit a URL, ask questions, get cited answers.
