MCP SERVERHand your agent a 50MB .cpuprofile and get a ten-line answer about which function is burning the CPU.
LABSCO SUMMARY
A production .cpuprofile is millions of lines of tick counts and call frames — reading it directly fills the context window and still needs aggregation across the call tree to mean anything. This does the aggregation outside the model and returns the conclusion, including the before-and-after diff that tells you whether your fix worked.
An MCP server that decodes Node.js V8 CPU profiles locally and returns compact bottleneck summaries instead of raw profile JSON.
A .cpuprofile file on disk — the one Node writes when you start it with CPU profiling turned on — and the path to it. No account and no key; the profile is decoded locally rather than uploaded.
One command — npx -y v8-cpu-profile-decoder-mcp
USE CASES
Rank functions by self time with V8 internals and Node built-ins filtered out, so the list is your code.
List every caller of a named function with invocation counts, which usually explains the hot path better than the function itself.
Map the hot functions through source maps to their original TypeScript file and line.
Diff the profile from before the change against the one after and read the per-function deltas.
Check GC time as a share of the profiling window before rewriting logic that is not the bottleneck.
Surface the V8 internal frames that stand for async machinery, which is where event-loop overhead shows up.
Counted once with only startup setup. Other versions, modes or optional credentials may expose a different list.
analyze_async_bottlenecksDetects event-loop overhead in a V8 CPU profile by identifying V8 internal frames that represent async machinery — microtask queue processing, nextTick saturation, and timer/immediate callbacks. These frames are invisible to most profilers but consume real CPU when promise chains are deep or nextTick is overused. Use to answer: is the bottleneck async orchestration overhead rather than synchronous computation?
analyze_call_tree_pathFinds all callers of a specific function in a V8 CPU profile and returns how often each caller invoked it. Accepts partial, case-insensitive function name matching. Use to answer: what is calling my slow function and how many times?
analyze_gc_pressureAnalyses a V8 .cpuprofile for garbage collection overhead. Reports total GC time as a percentage of profiling duration, broken down by GC type (Scavenger = short-lived object pressure, Mark-Sweep/Mark-Compact = old-space pressure, Incremental = high allocation rate). Flags when GC exceeds a configurable threshold and provides a targeted recommendation. Use to answer: is GC the bottleneck, and what kind of allocation pattern is causing it?
correlate_source_codeMaps the hottest functions in a V8 CPU profile back to their original TypeScript source locations using source map files (.js.map). Falls back to compiled JS locations if no source map is found. Use to answer: which TypeScript file and line is the bottleneck actually coming from?
diff_profilesCompares two V8 .cpuprofile files (before and after an optimization) and returns per-function CPU time deltas, normalized against each profile's total duration. Frames are matched by call-frame coordinates (functionName + url + line + column), not by transient node IDs, so alignment is stable across profiling sessions. Use to answer: which functions improved or regressed after my change, and by how much?
extract_hottest_functionsParses a V8 .cpuprofile file and returns the top N functions ranked by exclusive CPU time (self time). Filters out V8 internals and Node.js built-ins by default, returning only user code. Framework frames (express, next.js, koa, etc.) can be collapsed into a single entry. Recursive calls to the same source location are merged with an instanceCount field. Use this first to identify which functions are consuming the most CPU in a Node.js performance profile.
BEFORE YOU INSTALL
6 tools when we connected.
npx on PATH
CONTEXT COST
Rated by the size of the 6 tool definitions that ride along on every turn, not by how many there are
INSTALLATION
Add the block to your client, or run the command directly.
{
"mcpServers": {
"v8-cpu-profile-decoder-mcp": {
"command": "npx",
"args": [
"-y",
"v8-cpu-profile-decoder-mcp"
]
}
}
}npx -y v8-cpu-profile-decoder-mcpMCP servers
ZephyrDeng
ChromeDevTools
antarikshc
patchorbit