Built for Claude agents · MCP native · save your tokens

The database of
proven speedups.

63,518 performance tricks and pitfalls mined from 1,822 repositories and 2,294 scientific papers — each with its measured win and the commit that proves it. Built for Claude: one MCP call instead of a research loop.

every card links to its commit or paper88,892 mined → 63,518 kept by review10 free searches a day

Claude Codeclaude.ai connectorClaude Agent SDKany MCP clientx402

Real cards from the database

torvalds/linux 2.8× Split CRC into four independent streams, merged with PMULL folding constants proof: commit → FFmpeg/FFmpeg 6.2× Unroll an IIR recurrence by 4 with precomputed powers so outputs go independent proof: commit → duckdb/duckdb 10× Sliding-window quantiles reuse the prior partition instead of re-sorting each step proof: commit → llvm/llvm-project 37.5× Encode the register in the VarLoc ID so kill sets become interval lookups proof: commit →

Today

Every performance question starts a research loop.

Your agent searches, opens pages, reads repos and papers, compares claims and guesses which one held up. It pays for every page it reads — and pays again next time it forgets.

We did it once

We read it so your agent doesn't have to.

A local pipeline read every commit, hotspot file, pull request and paper passage, and kept only what has evidence: the bottleneck, the technique, the measured speedup, the trade-offs and when it applies. A reviewer model dropped the rest.

0repositories mined
0scientific papers
0candidate tricks extracted
0reviewed and kept
0pitfalls: what went wrong

The vector database

Every trick is a point in meaning space.

Each card is embedded as a 1024-dimension vector. This is a real 2D projection of of them, colored by domain. Similar mechanisms land close together, whatever the field.

    One query

    Ask once. The camera does the searching.

    Query:

    The index pulls its nearest neighbours out of 64,264 vectors in milliseconds — a video codec, a vector database, a JPEG library, a mesh optimizer. One mechanism, four fields. A web search never puts them side by side.

    Plus keywords and a graph

    Vectors, BM25 and links.

    Dense search finds the same idea in other words, BM25 finds the exact term, and the graph adds what each trick builds on, combines with, conflicts with, or was reverted by.

      Built for Claude

      Your agent asks. The database answers.

      adtomic is an MCP server. Claude Code, claude.ai and any MCP client call it like any other tool — signed in with OAuth or an API key, spending credits per call. Agents with no account pay per call over x402 and are only charged when the call succeeds.

      17 read-only toolsOAuth 2.1API keysHTTP 402 · USDC

      Save money and time

      Put in your own numbers.

      The defaults are examples, not measurements. One adtomic answer with 8 hits is about tokens, from the average size of a real card.

      Research loop
      adtomic (Pro)
      You save each month

      Connect

      Three ways in.

      Claude Code

      claude mcp add --transport http adtomic /mcp

      Run /mcp to sign in. CI and headless machines use an API key from your account page.

      claude.ai & Desktop

      Settings → Connectors → Add custom connector, URL /mcp. Sign in, allow, done — Claude searches adtomic before the web.

      Any agent, no account

      Call /x402/mcp. The reply quotes the price; the agent pays $0.01 per credit in USDC and retries. A failed call is never charged.

      Full setup guide Pricing

      One line for your CLAUDE.md: “Before researching a performance question on the web, query the adtomic MCP and cite the source URL.”