CPU Performance Engineering MCP Server: Setup for Claude, Cursor and Other Apps

• Finds the best evidence for a CPU performance question, with page numbers

  Searches the reading list, the benchmarks and the text of the linked papers and manuals, then returns the best passages with their sources. You can paste perf stat output, compiler remarks or assembly alongside the question and it reads the numbers for you.

  Try: “Here is my perf stat output. Where is the time going?”

  Read · ask

• Searches the list, its notes, its benchmarks or the linked sources

  Ranks matches for a few words across entries, sections, benchmarks, rejected candidates, notes and benchmark code. Switch the scope to sources to search inside the linked papers and manuals instead.

  Try: “Look up everything about store forwarding in the linked sources.”

  Read · lookup

• Finds documents by keyword and returns ids you can open

  A plain document search that returns titles, ids and links for entries, sections, benchmarks and source passages. It is the search tool ChatGPT deep research looks for.

  Try: “Search for documents about false sharing.”

  Read · search

• Opens one document found by search in full

  Takes an id from search, such as an entry or a source passage, and returns its full text with a link you can cite. Passage text comes from outside sources and is marked as untrusted.

  Try: “Fetch the document entry:4.3.5.”

  Read · fetch

• Shows the table of contents, or one section in reading order

  With no argument it returns the map of the 16 sections. Give it a number or title and it returns that section's entries in dependency order, with its benchmarks and, for the watchlist, what would promote each item.

  Try: “Show me the NUMA section.”

  Read · get_section

• Explains why one source is listed and what to read around it

  Returns one listed source with the reason it earned a place, what to read before and after, other places it appears, and any numbers or link notes tied to it. It also says whether the source text is in your local library.

  Try: “Tell me why Drepper's memory paper is on the list and what to read next.”

  Read · get_entry

• Gives an ordered reading plan for a topic

  Builds a study order from the list for a topic such as NUMA or branch prediction, ending with the benchmark that reproduces it. With no topic it returns the list's own starting path.

  Try: “Give me a reading path for branch prediction, ending with something I can run.”

  Read · reading_path

• Shows a runnable benchmark: claim, machine, results and code

  Returns one of the repository's C benchmarks with the claim it tests, the machine it ran on, results, analysis, limits, source code and raw output. With no name it lists them all.

  Try: “Show me the false sharing benchmark and how to run it.”

  Read · get_benchmark

• Shows why a source or number was kept out of the list

  Looks up rejected candidates and the rule each failed, performance claims with their verdicts, and link-check notes. Filter by rule, verdict or section.

  Try: “Why is cppreference not in the list?”

  Read · editorial_record

• Checks a performance claim against the list's seven-field rule

  Compares a claim with the rule that a number counts only with CPU model, core count, frequency and SMT state, compiler flags, workload, baseline and method. It says what is missing and how the list judged similar numbers. The check is heuristic.

  Try: “Is "AVX-512 gives 2x on Zen 5" a claim the list would quote?”

  Read · check_evidence

• Reads the text of one linked paper, manual or page

  Returns the text of a source the list links, from your local library or fetched live if it is not indexed yet. You can ask for matching passages, a PDF page or one passage in full.

  Try: “Read what the Intel optimisation manual says about store forwarding.”

  Network · read_source

• Opens a file from the repository bundled with the server

  Returns a repository file the server carries, such as the README, notes, benchmark source, build and run scripts or raw results. With no path it lists the available files.

  Try: “Show me the build script for the cache latency benchmark.”

  Read · read_file

• Shows how much of the linked material is indexed

  Reports which copy of the list is being served, whether the daily update worked, how many sources are indexed, and which are blocked and why. Ask for detail to see every source.

  Try: “How much of the source library has been indexed yet?”

  Read · library_status

Watch out for

• Builds a library on your disk on first run

  It downloads the linked sources in the background and stores their text, which can take up to a quarter of an hour and a few hundred megabytes. Set the data folder with CPU_PERF_DATA_DIR or turn this off with CPU_PERF_AUTO_INDEX=0. github.com

• Fetches web pages and updates itself daily

  It reads only URLs that appear in the list, over public addresses, and checks GitHub once a day for a newer list. Set CPU_PERF_AUTO_UPDATE=0 to stop the daily check and CPU_PERF_LIVE_FETCH=0 to stop live reads. github.com

• Source passages are untrusted text

  Passages come from outside documents and could contain instructions aimed at the assistant. The server labels them as untrusted data, but treat unexpected behaviour after a long answer with suspicion. github.com

• Optional HTTP mode has no sign-in

  Your client normally starts the server locally. If you run the shared HTTP instance, anyone with its address can call the tools, which only read. github.com

Security advisories