perfetto-mcp-rs
perfetto-mcp-rs
English | 简体中文
一个用于通过 LLM 分析 Perfetto 追踪文件的 MCP 服务器。将 Claude Code(或任何 MCP 客户端)指向 Perfetto 追踪文件(.pftrace / .perfetto-trace / .bin 等 —— 支持内容嗅探),并使用 PerfettoSQL 进行查询。
由 trace_processor_shell 提供支持 —— 首次运行时自动下载,无需手动安装 Perfetto。
最适合与支持多轮工具调用的智能体 MCP 客户端(Claude Code、Codex、Claude Desktop、Cursor)配合使用。非智能体客户端虽然也能看到相同的工具,但无法遵循引导 LLM 完成典型 load_trace → list_tables → list_table_structure → execute_sql 流程的错误消息提示。
引导智能体使用正确的 PerfettoSQL 标准库模块 —— 分析 SQL 始终由智能体自行生成。
快速安装
macOS / Linux (Homebrew, 推荐):
brew tap tooluse-labs/tap
brew install perfetto-mcp-rs
# brew prints caveats; run the printed line to register with Claude Code / Codex:
perfetto-mcp-rs install --binary-path "$(brew --prefix)/bin/perfetto-mcp-rs"Rust 开发者 (cargo install):
cargo install --locked perfetto-mcp-rs
perfetto-mcp-rs install --binary-path "$(which perfetto-mcp-rs)"Linux / macOS / Windows (Git Bash, MSYS2, Cygwin) — 直接脚本:
curl -fsSL https://raw.githubusercontent.com/tooluse-labs/perfetto-mcp-rs/main/install.sh | shWindows (PowerShell):
irm https://raw.githubusercontent.com/tooluse-labs/perfetto-mcp-rs/main/install.ps1 | iex上述安装程序会将预构建的二进制文件放入 ~/.local/bin(Windows 上为 %USERPROFILE%\.local\bin),并在需要时将其添加到您的用户 PATH 中。如果安装了 Claude Code 和/或 Codex,它还会自动进行注册。重启 Claude Code 或开启新的 Codex 会话即可生效。
Claude 作用域: 默认注册为 --scope user(可在任何目录下使用)。如需项目级安装,请设置 SCOPE=local(或 project)并在该项目目录下运行脚本:
SCOPE=local bash -c 'curl -fsSL https://raw.githubusercontent.com/tooluse-labs/perfetto-mcp-rs/main/install.sh | sh'PowerShell 等效命令:$env:SCOPE = 'local'; irm ... | iex。Codex 没有作用域概念,会忽略此变量。
支持的平台:linux amd64/arm64, macOS amd64/arm64, Windows amd64。如果您不想运行脚本,可以直接从 发布页面 获取二进制文件。发布资源命名为 perfetto-mcp-rs-<platform>(例如 perfetto-mcp-rs-linux-amd64);在调用 install 时重命名或显式指定下载的文件,并在 Unix 系统上先标记为可执行文件 (chmod +x) —— 子命令会拒绝非可执行路径,以避免写入损坏的 MCP 条目。示例:
# Linux amd64 example — adjust the asset name for your platform.
curl -fsSL -o perfetto-mcp-rs \
https://github.com/tooluse-labs/perfetto-mcp-rs/releases/latest/download/perfetto-mcp-rs-linux-amd64
chmod +x perfetto-mcp-rs
./perfetto-mcp-rs install --scope user --binary-path "$PWD/perfetto-mcp-rs"Related MCP server: mcp-server-logs-sieve
升级
重新运行相同的安装命令 —— 它会拉取最新版本,安全地覆盖现有二进制文件(支持 Windows 文件锁定重试),并幂等地重新向 Claude Code / Codex 注册 MCP 服务器。
使用 --version 标志锁定特定版本:
curl -fsSL https://raw.githubusercontent.com/tooluse-labs/perfetto-mcp-rs/main/install.sh | sh -s -- --version v0.7.0VERSION 环境变量也有效,但必须紧接在 sh 之前(POSIX VAR=value cmd 仅作用于下一条命令 —— VERSION=v0.7.0 curl ... | sh 会将 VERSION 作用于 curl,而不是管道后的 sh):
curl -fsSL https://raw.githubusercontent.com/tooluse-labs/perfetto-mcp-rs/main/install.sh | VERSION=v0.7.0 shPowerShell —— 在同一行设置 $env:VERSION,因为 iex 在当前会话中运行:
$env:VERSION = 'v0.7.0'; irm https://raw.githubusercontent.com/tooluse-labs/perfetto-mcp-rs/main/install.ps1 | iex没有自动更新守护进程 —— 升级是显式的。
检查更新
perfetto-mcp-rs check-update如果已是最新(或领先于发布版本 —— 本地开发构建),则退出码为 0;如果存在更新版本,则为 2;网络或解析错误则为 1。适用于 shell 提示符集成和 CI 预检查。
卸载
每个平台都有对称的单行命令。从 Claude Code 和 Codex 中注销,删除二进制文件,并删除缓存的 trace_processor_shell。幂等 —— 如果某些步骤已手动完成,再次运行也是安全的。
Linux / macOS / Windows (Git Bash, MSYS2, Cygwin):
curl -fsSL https://raw.githubusercontent.com/tooluse-labs/perfetto-mcp-rs/main/uninstall.sh | shWindows (PowerShell) — 请先关闭 Claude Code、Codex 或任何其他正在使用该 .exe 的程序:
irm https://raw.githubusercontent.com/tooluse-labs/perfetto-mcp-rs/main/uninstall.ps1 | iex作用域安装 (local / project): claude 按项目目录存储本地/项目条目,因此卸载必须使用相同的 SCOPE 并在该目录下运行。省略此步骤会留下作用域内的 Claude 条目,而包装器仍会删除二进制文件和缓存:
# Ran `SCOPE=local bash install.sh` in ~/work/foo earlier? Then:
cd ~/work/foo
SCOPE=local bash -c 'curl -fsSL https://raw.githubusercontent.com/tooluse-labs/perfetto-mcp-rs/main/uninstall.sh | sh'PowerShell 等效命令:cd <original-project-dir>; $env:SCOPE = 'local'; irm ... | iex。
$INSTALL_DIR(默认 ~/.local/bin)不会从您的 PATH 中移除:
Linux / macOS — 安装程序仅打印
PATH提示;如果您将其添加到了 shell rc 文件中,请手动删除该行。Windows — 安装程序将
$INSTALL_DIR写入您的用户 PATH (HKCU\Environment);如果您想删除它,请通过“系统属性”→“环境变量”进行操作。
其他工具可能仍依赖此目录,这就是卸载时将其保留的原因。
工具
工具 | 用途 |
| 打开 Perfetto 追踪文件(任何 trace_processor 可读格式: |
| 列出已加载追踪中的表/视图,可选 GLOB 过滤器 |
| 显示表的列名和类型 |
| 运行 PerfettoSQL 查询,返回列式 JSON |
| 列出追踪中的进程(pid, 名称, 开始/结束时间戳) |
| 列出进程名称下的线程(最多 2000 个) |
| 最严重的卡顿帧及其原因、子原因、delay_since_last_frame(Chrome 追踪) |
| 页面加载:URL, FCP, LCP, DCL, 加载耗时(毫秒)(Chrome 追踪) |
| 按持续时间排序的主线程任务(含 cpu_pct,使用 is_main_thread)(Chrome 追踪) |
| 浏览器启动事件和首次可见内容时间(Chrome 追踪) |
| 按持续时间排名的 Web 内容交互(点击、轻触、INP)(Chrome 追踪) |
| 列出可用的 PerfettoSQL 标准库模块及使用示例(无需追踪文件) |
典型流程取决于追踪类型:
Chrome 追踪:
load_trace→ 专用chrome_*工具(chrome_scroll_jank_summary,chrome_page_load_summary,chrome_main_thread_hotspots,chrome_startup_summary,chrome_web_content_interactions) →execute_sql以对返回的行进行更深入的分析。其他追踪:
load_trace→list_tables/list_table_structure以进行模式发现 →execute_sql进行查询。当标准库模块可能涵盖您的分析时(Android,通用模块如slices.with_context),可辅助调用list_stdlib_modules。
示例
向 Claude Code 或 Codex 询问类似以下内容:
加载
~/traces/scroll_jank.pftrace并告诉我最主要的滚动卡顿原因。
Claude 将调用 load_trace,然后发出类似以下的查询:
INCLUDE PERFETTO MODULE chrome.scroll_jank.scroll_jank_v3;
SELECT cause_of_jank, COUNT(*) AS n
FROM chrome_janky_frames
GROUP BY cause_of_jank
ORDER BY n DESC;手动 MCP 客户端配置
如果安装程序的自动注册不适用于您的客户端:
Codex:
codex mcp add perfetto-mcp-rs -- /absolute/path/to/perfetto-mcp-rs基于 JSON 的客户端 (例如 Claude Code, Claude Desktop, Cursor):
{
"mcpServers": {
"perfetto-mcp-rs": {
"command": "/absolute/path/to/perfetto-mcp-rs"
}
}
}配置
变量 | 作用 |
| 指向现有 |
| 覆盖 |
| 覆盖 HTTP 状态/查询超时时间(毫秒) |
|
|
CLI 标志:
标志 | 默认值 | 描述 |
| 3 | 最大缓存的 |
| 20000 | 等待生成的 |
| 30000 |
|
从源码构建
需要 Rust 工具链和 protoc(Protocol Buffers 编译器):
# Ubuntu/Debian
sudo apt install -y protobuf-compiler
# macOS
brew install protobuf
# Windows
choco install protoc然后:
git clone https://github.com/tooluse-labs/perfetto-mcp-rs
cd perfetto-mcp-rs
cargo build --release
# Binary at target/release/perfetto-mcp-rs开发
cargo test # unit tests
cargo clippy # lint
cargo fmt # format许可证
采用 Apache License, Version 2.0 或 MIT license 双重许可。贡献均在相同条款下接受。
Available Tools
17 toolschrome_main_thread_hotspotsARead-onlyIdempotent
Top Chrome main-thread tasks by wall duration: id, ts, name, task_type, thread_name, process_name, upid, pid, nullable machine_id, dur_ms, overlap_dur_ms, full_task_cpu_pct/full_task_thread_dur_ms, overlap_cpu_pct/overlap_thread_dur_ms; legacy cpu_pct/thread_dur_ms are full-task. Uses chrome.tasks, thread.is_main_thread = 1, and Chrome's Cr*Main fallback. Pass a returned id to slice_descendants_breakdown for child-slice breakdowns.
Use when: investigating responsiveness, scroll/load stalls, CPU vs wall time, or one renderer.
Don't use for: non-Chrome traces (will error). For background (non-main) thread tasks, drop to execute_sql against chrome.tasks directly.
Parameters (all optional):
process_name/pid/machine_id/upid: scope to one process/type. Preferupid; addmachine_idto disambiguate multi-machinepids. All filters AND.page_load_id/navigation_id/phase: scope to a page-load window. IDs matchchrome_page_loads.idand.navigation_idrespectively and are mutually exclusive.phase:navigation_to_fcp,navigation_to_load,dcl_to_fcp,fcp_to_load. If an id is set withoutphase, defaults tonavigation_to_fcp; phase-only uses the latest page load.start_ts_ns/end_ts_ns: raw trace timestamp bounds in nanoseconds (end_ts_nsexclusive); aliasesstart_ts/end_tsare accepted; intersect page-load windows.overlap_dur_msis clipped to that window.min_dur_ms: minimum full-task duration, or clipped overlap duration when a window is set. Defaults to 16 ms. Pass 0 for all positive-overlap tasks.limit: max rows (default 100, capped at 5000). Must be > 0 if set.max_string_len: optional cap for returned string cells. Unset preserves full strings for precision. Must be > 0 if set.
Output: metadata-first JSON preserving columns / rows; truncated=true means an extra-row probe found more rows; string_truncated=true means cell text was shortened.
Empty result: no detected main-thread tasks exceeded min_dur_ms at the selected process/window threshold, or the trace uses non-standard main-thread names outside the Cr*Main fallback.
Inputs: pid, upid, limit, phase, +9 more.
chrome_page_load_resource_hotspotsARead-onlyIdempotent
Rank URL-bearing Chrome resource/request slices in a page-load/raw window. Returns timing, overlap, process/thread/machine_id, URL. Use after chrome_page_load_resource_summary to drill into slow URL slices. Filters: page_load/window, min_dur_ms default 50, limit, max_string_len. slice_duration_status='incomplete_duration' means dur=-1; overlap is measured to window end or trace_end().
Inputs: limit, phase, trace_id, end_ts_ns, +5 more.
chrome_page_load_resource_pipelineARead-onlyIdempotent
Drill into one Chrome page-load resource URL and join lifecycle/request spans with script/style/layout signals. Use after chrome_page_load_resource_summary with example_slice_id or url_substring. Returns timing facts, match evidence, resource ends, primary/detail slice evidence (detail is not a root-cause claim), URL/ancestor-correlated resource result source/hit, machine-id sets, incomplete count, and evidence_boundary. Parameters: URL seed required; optional window filters, url_grouping, limit default 30, max_string_len.
Inputs: limit, phase, trace_id, end_ts_ns, +7 more.
chrome_page_load_resource_summaryARead-onlyIdempotent
URL-level Chrome resource/request summary for a page-load/raw window. Returns URL key, process/machine/priority sets, span, max/summed overlap, primary/detail slice evidence (detail is not a root-cause claim), navigation/renderer relation including target_renderer_source, example_slice_id, incomplete_duration_slice_count. Use before chrome_page_load_resource_hotspots; rank by max overlap.
Inputs: limit, phase, trace_id, end_ts_ns, +6 more.
chrome_page_load_script_hotspotsARead-onlyIdempotent
Rank renderer main-thread script groups in a Chrome page-load/raw window: URL/name/process/thread/machine_id, wall/CPU totals, style/layout ms, example_slice_id. Read-only.
Use when: slow FCP/load needs post-resource JS attribution; expand example_slice_id with slice_descendants_breakdown.
Parameters: optional process filters (process_name/pid/machine_id/upid), page-load/window filters shared with chrome_main_thread_hotspots, min_total_ms (default 20), limit, max_string_len. Empty result: no matching script groups.
Inputs: pid, upid, limit, phase, +9 more.
chrome_page_load_summaryARead-onlyIdempotent
Summarize each page navigation in a Chrome trace: navigation id, URL, raw boundary timestamps, FCP / LCP / DCL / load timings in ms. Read-only.
Use when: comparing page-load timings across navigations, finding slow loads, baselining web-vitals before/after a change. Prefer over hand-joining chrome.page_loads — schema is already correct.
Don't use for: non-Chrome traces (will error). For sub-event timings inside one navigation, drop to execute_sql against the chrome.page_loads module.
Parameters: optional limit (default 100, capped at 5000) and max_string_len. Operates on the loaded trace.
Output: metadata-first JSON; row_count exact; truncated=true means more rows exist; string_truncated=true means shortened text.
Empty result: no navigations occurred during capture (e.g. trace started after the page was already loaded).
Inputs: limit, trace_id, max_string_len.
chrome_scroll_jank_summaryARead-onlyIdempotent
Summarize the worst scroll jank frames in a Chrome trace: cause_of_jank, sub_cause_of_jank, delay_since_last_frame, event_latency_id, scroll_id, vsync_interval. One row per janky frame, sorted by delay_since_last_frame DESC. Read-only.
Use when: investigating jank reports, finding scroll regressions, ranking jank causes. Prefer over hand-rolling SQL on chrome.scroll_jank.scroll_jank_v3 — same data, less code.
Don't use for: non-Chrome traces (will error). For custom filters, use execute_sql against the same view.
Parameters: optional limit (default 100, capped at 5000) and max_string_len. Operates on the loaded trace.
Output: metadata-first JSON; row_count exact; truncated=true means more rows exist; string_truncated=true means shortened text.
Empty result: no janky frames detected (clean trace) or no scrolls occurred during capture.
Inputs: limit, trace_id, max_string_len.
chrome_startup_summaryARead-onlyIdempotent
Summarize Chrome browser startup events: id, name, launch_cause, startup_duration_ms (first_visible_content_ts - startup_begin_ts), browser_upid. Read-only.
Use when: measuring time-to-first-visible-content for cold starts, comparing launch causes (NEW_WINDOW vs CMD_LINE vs RESTORE_SESSION), regressing startup performance.
Don't use for: non-Chrome traces (will error). Browser-process work during steady state is covered by chrome_main_thread_hotspots.
Parameters: optional limit (default 100, capped at 5000) and max_string_len. Operates on the loaded trace.
Output: metadata-first JSON; row_count exact; truncated=true means more rows exist; string_truncated=true means shortened text.
Empty result: trace started after the browser was already running (most cases — startup is captured only when tracing began before launch).
Inputs: limit, trace_id, max_string_len.
chrome_web_content_interactionsARead-onlyIdempotent
Rank Chrome web content interactions by total_duration_ms: id, ts, total_duration_ms, longest_event_dur_ms, interaction_type, renderer_upid. Read-only.
Use when: INP analysis, reproducing user-felt latency, finding slow click/tap/keyboard handlers.
Don't use for: non-Chrome traces (will error). For interactions filtered by interaction_type, drop to execute_sql against chrome.web_content_interactions.
Parameters: optional limit (default 100, capped at 5000) and max_string_len. Operates on the loaded trace.
Output: metadata-first JSON; row_count exact; truncated=true means more rows exist; string_truncated=true means shortened text.
Empty result: no interactions captured (trace started before user input or interaction tracking was disabled in tracing config).
Inputs: limit, trace_id, max_string_len.
execute_sqlARead-onlyIdempotent
Run a PerfettoSQL query against the loaded trace and return rows as columnar JSON. Read-only against trace data; SQLite operates in-memory per session. Aggregates are strongly preferred over raw row data; results are capped at 5000 rows.
Use when: composing analyses not covered by the dedicated tools — custom aggregations, joins across stdlib modules, or queries against base tables (slice, thread, process, sched).
Don't use for: questions the dedicated chrome_* tools answer — they return the same data with the JOIN shape already correct. Don't hand-roll slice scans with LIKE '%x%' patterns when a stdlib module covers the data; INCLUDE PERFETTO MODULE chrome.tasks is faster and the joins are pre-baked.
Parameters: sql is a single PerfettoSQL statement (the INCLUDE PERFETTO MODULE foo; and SELECT ... can be in the same call). Optional output shaping (head/limit, columns_only, summary, include_row_count, max_string_len) only changes what this tool returns; it does not rewrite the SQL. Blob cells render as blob:hex:<hex>. String results may be redacted by the server privacy policy before they are returned, preserving diagnostic structure while masking sensitive URL/header/cookie/path values. Requires load_trace to have run first.
Empty rows means the query matched nothing — distinct from a SQL error, which is returned as an error string with a hint pointing at the most likely cause (missing module, missing column, missing table).
Reference docs (fetch when you need exact column names or function signatures): https://perfetto.dev/docs/analysis/stdlib-docs (24 stdlib packages — chrome / android / sched / slices / linux / wattson / v8 / ...; use per-package anchors like #package-chrome), https://perfetto.dev/docs/analysis/perfetto-sql-syntax (syntax).
Inputs: sql (required), head, limit, summary, +4 more.
list_processesARead-onlyIdempotent
List every process captured in the trace: upid (trace-internal id), pid, machine_id, name, start_ts, end_ts. Read-only.
Use when: entry point for Android and Linux trace analysis, or picking the right pid/upid to feed into list_threads_in_process or chrome_main_thread_hotspots.
Don't use for: Chrome traces — the dedicated chrome_* tools answer most common questions without process-level navigation.
Parameters: none — operates on the loaded trace.
Empty result: rare; would mean the trace captured no process metadata at all.
Errors when: no trace is loaded — call load_trace first.
Inputs: trace_id.
list_stdlib_modulesARead-onlyIdempotent
List curated PerfettoSQL stdlib modules as JSON entries with domain, module, views, description, and usage. Use when choosing an INCLUDE PERFETTO MODULE ... target; no trace has to be loaded.
Optional filters: domain (chrome, android, generic), query (case-insensitive search over module/view/description), and limit. For longer guidance, read resource://perfetto-mcp/stdlib-quickref.
Inputs: limit, query, domain.
list_tablesARead-onlyIdempotent
List tables and views in the loaded trace. Read-only.
Use when: exploring an unfamiliar trace or verifying a table exists before writing SQL. Underlying SQL engine is SQLite, so the catalog tables common in other SQL engines aren't present — this MCP tool is the schema introspection path.
Don't use for: queries against known stdlib modules — go straight to execute_sql with INCLUDE PERFETTO MODULE. Don't reference this tool name inside SQL; it's a separate MCP tool, not a SQL function — call it via the tool API.
Parameters: optional pattern — SQLite GLOB filter (e.g. chrome_* for chrome stdlib views, slice* for the slice table family). Without it, internal stdlib tables (_*) are hidden.
Empty result: no tables matched. For stdlib views, run execute_sql with INCLUDE PERFETTO MODULE ... first; otherwise retry an explicit pattern for internal tables.
Errors when: no trace is loaded — call load_trace first.
Inputs: pattern, trace_id.
list_table_structureARead-onlyIdempotent
Show the columns of a table or view: name, type, nullability, primary_key flag.
Use when: writing or debugging a query — call this immediately after a no such column error to inspect the actual schema rather than guessing. Both stdlib views and base tables have fixed schemas; don't infer columns by analogy across them.
Don't use for: this is a separate MCP tool, not a SQL function — don't write SELECT * FROM list_table_structure inside execute_sql.
Parameters: table_name (string) — the exact table or view name as it appears in list_tables output. Case-sensitive; does not accept GLOB patterns or partial matches. Also accepts the alias name (v0.11.3+).
Errors when: the table doesn't exist or has no columns. Call list_tables first; stdlib views may need an INCLUDE first.
Inputs: table_name (required), trace_id.
list_threads_in_processARead-onlyIdempotent
List threads in one process or same-named process set: tid, thread_name, pid, upid, machine_id. Limit 2000, cap 5000.
Use when: drilling into a process from list_processes.
Don't use for: ALL trace threads — use execute_sql on thread.
Parameters: pass either upid (trace-internal id, precise — prefer when multiple processes share a name like 'Renderer') or process_name (exact match). upid wins when both are set. Optional limit and offset page large result sets; both accept numbers or numeric strings.
Output: exact row_count, returned_rows, truncated/has_more; rows are ordered by pid/tid. process_counts reports per-upid counts for same-name fan-out.
Empty result: returned as an error pointing at list_processes for available candidates.
When truncated=true, increase offset or drill down with upid.
Inputs: upid, limit, offset, trace_id, +1 more.
load_traceAIdempotent
Load one or more local Perfetto trace files and return routing summaries plus trace_id handles. Other tools use the active trace by default; pass trace_id to analyze a specific loaded trace after loading more.
Use when: starting any analysis session — call this first.
Don't use for: live capture or streaming URLs; paths must be complete local trace files.
Parameters: pass path for one trace or paths for a batch. Calling again makes the last loaded path the active default; cached trace processors make repeat loads cheap unless the file fingerprint changed.
Errors when: the file doesn't exist, isn't a valid Perfetto trace, trace_processor fails to parse it, or the binary cannot be downloaded.
Inputs: path, paths.
slice_descendants_breakdownARead-onlyIdempotent
Recursive child-slice expansion under known slice.id roots, aggregated as a bounded breakdown per (depth, name) group. Use to drill into a long task — after chrome_main_thread_hotspots or execute_sql returns a slice id — without hand-writing WITH RECURSIVE CTEs over slice.parent_id. Required: slice_ids. Optional bounds: min_dur_ms, max_depth, limit, include_args, max_string_len. The response echoes summary_scope, applied_filters, and missing_root_ids (missing root slice ids). Returned columns: root_id, depth, name, slice_count, inclusive_total_ms (do not sum across depths), self_ms (direct-child time subtracted, clamped at zero), max_ms, first_ts_ns (raw nanoseconds, not ms), example_slice_id (longest-duration descendant per group), and optionally example_args. incomplete_descendant_count counts dur<0 descendants excluded from duration aggregates.
Inputs: slice_ids (required), limit, trace_id, max_depth, +3 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 17 tools
Most tools have distinct purposes (page-load summaries, resource hotspots, main-thread hotspots, scroll jank, startup, interactions). However, `chrome_page_load_resource_summary`, `chrome_page_load_resource_hotspots`, and `chrome_page_load_resource_pipeline` all deal with resource URL analysis and could be confused, though their descriptions clarify the intended sequence (summary → hotspots → pipeline). The generic `execute_sql` and `list_*` tools are clearly separate.
All tools follow a consistent snake_case pattern with a clear domain prefix: `chrome_*` for Chrome-specific analysis, `list_*` for exploration, `execute_sql`, `load_trace`, `slice_descendants_breakdown`. Verbs are uniform (list, load, execute, chrome_page_load_*, chrome_main_thread_*) and the naming clearly indicates the tool's function.
17 tools is on the higher side but justified for a comprehensive Perfetto trace analysis server covering Chrome performance domains (page loads, resources, main thread, jank, startup, interactions) plus general SQL and schema exploration. Each tool serves a distinct purpose, and the count reflects the breadth of the domain without being bloated.
The tool set covers major Chrome performance analysis workflows (page loads, resources, main-thread tasks, scroll jank, startup, web interactions) plus generic trace exploration (processes, threads, tables, SQL). Minor gaps exist: no dedicated tool for memory or CPU profile analysis, and the resource pipeline depends on the summary/hotspots sequence, but `execute_sql` covers custom needs.
Maintenance
Related MCP Connectors
Shared memory MCP server for Claude Code, Codex, Gemini CLI and Cursor. Cited, editable.
A comprehensive Model Context Protocol (MCP) server that enables AI assistants to interact with yo…
MCP server for progressive tool usage at any scale (see https://klavis.ai)
Related MCP Servers
- FlicenseAqualityDmaintenanceA FastMCP server that provides LLMs with structured access to Scalene's CPU, GPU, and memory profiling for Python applications. It enables automated performance analysis, bottleneck identification, and optimization suggestions through natural language interactions in supported IDEs.71-
- AlicenseNot gradedqualityBmaintenanceAn MCP server that connects Claude (or any MCP compatible client) to your existing log infrastructure. Query, summarize, and trace logs in plain English across GCP Cloud Logging, AWS CloudWatch, Azure Log Analytics, Grafana Loki, and Elasticsearch without writing filter expressions or leaving your editor.5MIT
- AlicenseNot gradedqualityBmaintenanceAn MCP server that converts Windows WPR .etl performance traces into structured JSON summaries and flamegraph-ready data for LLM analysis. It bridges Windows Performance Analyzer automation with LLM reasoning capabilities for performance troubleshooting.1MIT
- AlicenseAqualityDmaintenanceMCP server that gives AI agents access to your application's OpenTelemetry traces for querying, analysis, and debugging.511 npm2MIT