SuperAPI AI Directory

cpp26-adapter

C++26 paper reference backing a Claude Code plugin: paper lookup, fuzzy search and compiler support status over the ISO/IEC 14882:2026 corpus.

Open website

About this tool

cpp26-adapter is a Claude Code plugin that can also be used as a standalone MCP server. Its goal is to turn LLM coding assistants into C++26 experts, biasing generated results toward the idioms of the final form of ISO/IEC 14882:2026.

  1. Main capabilities The plugin includes a complete set of components: skill, MCP, reviewer, hooks, and slash commands. Its core capabilities include:

  2. Generate idiomatic C++26 code: bias the model toward C++26 idioms before generation, so that without repeated prompting you get pre(...) instead of assert(...), std::execution instead of std::async, and reflection instead of X-macros.

  3. Review existing code for C++26 modernization: run @cpp26-reviewer src/foo.cpp, and the subagent performs a two-pass review—first a regex anti-pattern check, then runs clang -std=c++2c -fsyntax-only, classifying each diagnostic as either compiler lag (the paper has been adopted, but the compiler has not yet implemented it) or a bug (a real code error).

  4. Initialize a new C++26 project: run /cpp26-init to generate CMakeLists.txt (CXX_STANDARD 26, -std=c++2c, CMAKE_EXPORT_COMPILE_COMMANDS ON), .clangd, and .cpp26-adapter.yaml.

  5. Retrieve authoritative paper citations: the lookup_paper MCP tool returns the normative text of a paper, making it easy to cite or to have the LLM answer based on the actual proposal rather than potentially outdated training data.

  6. How to use Full plugin installation: /plugin install cpp26-adapter@parasxos/claude-plugins. After installation, wiring is automatic: the skill loads context, MCP starts on demand, the review subagent can be invoked via @cpp26-reviewer, hooks lint on every C++ edit, and /cpp26-init is available. MCP server only (any MCP-compatible client): run pip install cpp26-ref on a reachable Python (≥3.11). Claude Desktop configuration: add the server to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows). The MCP server exposes three tools over stdio, based on an in-memory index of 216 C++26 papers (no SQLite, no embeddings, no build step). Cold start is about 352 ms, and 17 tests under pytest mcp-server/tests pass.

  7. Applicable scenarios Suitable for code generation and review where you want to directly adopt C++26 idioms, for example implementing "enum to string" directly with P2996 reflection + P1306 expansion statements, with no macros, third-party dependencies, or code generation; it is also suitable for scenarios that require citing the original text of papers, evaluating existing code for modernization, or quickly setting up a C++26 project skeleton.

  8. Limitations and notes It is recommended to follow the standard rather than the toolchain: even if clang < 22 / gcc < 16 has not yet released a feature, the plugin still recommends the C++26 form; compiler errors are only marked as compiler lag or bugs and are never automatically rewritten to pre-C++26 alternatives. The recommended path (skills + MCP) is compiler-independent; compiler awareness exists only in the reviewer's second-pass syntax check and the SessionStart probe. Evaluation data: in a 39-task suite, the plugin scores 37/39 (95%) when enabled and 14/39 (36%) when disabled, with a threshold of ≥85%. In v1.0.0, the evaluation threshold remained passing after two consecutive refreshes; the MCP tool signatures, subagent output schema, and standard-first invariants are binding for the 1.x series. License: the code (outside corpus/) is MIT, and the knowledge corpus (corpus/) is CC-BY-SA 4.0. Other details such as fees and pricing are not provided in the materials.