v0.1.7

search_docs responses can now carry an optional see_also field, a one-sentence pointer to list_targets that appears when the results are target-selection pages. Agents mid-task follow their latest results rather than the tool descriptions they read at the start, so the routing hint rides inside the response. In a routing eval on a multi-GPU vs multi-node target question, agents called list_targets 2 out of 3 times with the field and never without it.

No index schema change; existing indexes remain valid.

v0.1.6

Routing and input hardening, driven by a 12-question tool-routing eval.

• find_api's description now redirects target-option-value questions (for example what set_target accepts) to list_targets. This was the one failing question class in the eval; it reaches find_api because it names a symbol, so the redirect lives on find_api's side.
• category on list_targets is now enum-constrained to "simulator", "hardware", or "cloud". Invalid values fail schema validation with the legal values shown; a value that still reaches the handler returns an error naming them instead of a silent empty list.

No index schema change; existing indexes remain valid.

v0.1.5

Hardens link handling against code-shaped targets (#7).

• A relative link is absolutized only if its target has a path separator or a letter file extension; a lambda default like [](x=1.5) (a numeric literal, not a doc suffix) is left verbatim. No page in the current docs was affected; this prevents corruption of any future bare-indented C++ example.

Community project, not affiliated with or endorsed by NVIDIA. CUDA-Q is a trademark of NVIDIA Corporation.

v0.1.4

Fixes escaped-bracket corruption in prose (#5).

• Backslash-escaped brackets are no longer treated as leftover spans or as display math, so numbered specification clauses (**[1]**), type annotations ([float]), and attributes ([[nodiscard]]) render as literals. Genuine spans are still stripped and real math is preserved.

Community project, not affiliated with or endorsed by NVIDIA. CUDA-Q is a trademark of NVIDIA Corporation.

v0.1.3

Fixes documentation-fidelity bugs on the C++ API reference page (#3).

• Nested example code is now fenced. Highlight blocks nested in Doxygen definition lists were not recognized and flowed as prose; #include lines and lambda kernels are now proper code blocks.
• C++ signatures read as literal C++. Pandoc-escaped punctuation in prose (std::vector\<double\>) is reversed, while LaTeX math spans are preserved.

Community project, not affiliated with or endorsed by NVIDIA. CUDA-Q is a trademark of NVIDIA Corporation.

v0.1.2

Fixes a documentation-corruption bug.

• Code is no longer mangled by link rewriting (#1). C++ lambda parameter lists such as [](float theta) were being rewritten into absolute URLs, producing C++ that could not compile. Cleanup and link rewriting are now confined to prose, leaving fenced blocks, inline code, and indented code examples untouched.

Community project, not affiliated with or endorsed by NVIDIA. CUDA-Q is a trademark of NVIDIA Corporation.

v0.1.1

Adds MCP registry metadata: server.json and the registry ownership marker in the README. No functional changes.

v0.1.0

First release. An MCP server that serves NVIDIA CUDA-Q documentation, API reference, and runnable examples to AI agents, version-pinned to the installed cudaq package.

• Five tools: search_docs, get_page, find_api, search_examples, and list_targets
• Two resources: cudaq://versions and cudaq://llms.txt
• Offline SQLite FTS5 index built from CUDA-Q's published markdown mirrors and Sphinx inventory
• Install: pip install cudaq-docs-mcp (or uvx cudaq-docs-mcp)

Community project, not affiliated with or endorsed by NVIDIA. CUDA-Q is a trademark of NVIDIA Corporation.

Prebuilt docs indexes

Rolling release: nightly-refreshed prebuilt indexes downloaded by the server on first run.