perfonext
OPEN SOURCE. BUILT FOR NEXT.JS.
Performance evidence.
For your AI assistant.
CPU profiles, React renders, and build artifacts.
Three MCP servers that turn them into context your coding assistant can act on.
Claude Code
GitHub Copilot
Claude Desktop
Cursor
01 / THE EVIDENCE
Your assistant brings the reasoning.
perfonext brings the evidence.
Self time excludes callees. The 500 ms recording includes 100 ms of idle time.
02 / THE TOOLKIT
Use one. Use all three.
Each works independently.
01 / CPUV8 and Chrome CPU profiles, distilled into function hotspots, package costs, and caller relationships.
Source annotations depend on available sample data and files inside the server's working directory.
@perfonext/profiler-mcp
9 TOOLSRead the docs 
02 / RENDERSLive React captures or DevTools Profiler exports, distilled into component costs and rerender evidence.
Exact causes need captured change descriptions. Without them, findings are explicitly heuristic.
@perfonext/render-mcp
10 TOOLSRead the docs 
03 / BUILDNext.js build artifacts, distilled into route footprints, shared chunks, and the evidence behind bundle growth.
Import chains and package attribution need webpack stats. Manifest analysis works without them.
@perfonext/build-mcp
11 TOOLSRead the docs 03 / THE FEEDBACK LOOP
In the tools you already use.
With the context that was missing.
Capture a CPU or render profile, or point at a .next build. The servers provide collection recipes when you need them.
Your assistant calls the MCP tools and gets structured evidence: ranked costs, relationships, and comparisons.
MCP TOOLS → STRUCTURED RESULTSUse that evidence to guide a code change. Capture the same scenario again and compare what changed.
YOUR ASSISTANT → YOUR CODEThe servers provide the evidence. Your coding assistant proposes and applies the changes.
04 / GET CONNECTED
Node.js and npm required.
No global install needed.
Standard MCP over stdio.
Runs locally. MIT licensed.
claude mcp add perfonext-profiler -- npx -y @perfonext/profiler-mcp
claude mcp add perfonext-render -- npx -y @perfonext/render-mcp
claude mcp add perfonext-build -- npx -y @perfonext/build-mcpRun in your project's terminal, then approve the servers in Claude Code.
“How do I capture a CPU profile of my Next.js server?”
Using nvm on macOS?
GUI apps launched from the Dock may not see your Node installation. For spawn npx ENOENT, set an absolute npx path and include the same Node bin directory in the server's env.PATH.
BEFORE YOU DIVE IN
No. perfonext is an agent companion to your profiling tools. It turns profiles and build artifacts into structured summaries, source-aware follow-up, and comparisons that an AI coding assistant can work with. You can still use DevTools to inspect the original data.
The live path uses react-scan/lite and a local ingest endpoint. Instrumentation must run before React initializes. Use next dev, or next build --profile followed by next start; a plain production build has no profiling hooks.
Development preserves component names but adds timing overhead. Production profiling has more representative timings, but names can be minified. Capture also works with headless Playwright tests; a browser extension is not required.
The manifest-based tools don't need webpack stats. Import chains, duplicate-package attribution, and shared-chunk package breakdowns do. Turbopack doesn't provide that webpack module graph; how_to_collect_stats explains the available paths.
The MCP servers analyze your artifacts locally and send tool results to your configured MCP client. There is no perfonext backend. Your AI client's own data handling and model-provider settings still apply to the results it receives, including source context you request.