quickai

Profiler for Claude Code. See exactly where your tokens, money and time go — per task, per subagent, per model — straight from the transcripts Claude Code already writes to disk. No instrumentation, no telemetry, nothing leaves your machine.

Most Claude Code users are on a flat subscription, so quickai leads with tokens (and time), not dollars. The $ figures are an API-equivalent (what those tokens would cost pay-as-you-go) — a measure of volume and ROI, not a bill.

• Where the money/tokens went — top tasks, projects, models, agent types
• Where the time went — wall vs idle, per-task timeline, "why was the harness slow?"
• Behavior profiling — tool usage & error rates, cache-hit health, spinning tasks, waste
• Per-project reports and a self-contained HTML report you open in the browser
• MCP server — ask all of the above straight from a chat

1. Installation

Homebrew (macOS / Linux)

brew install AlexGladkov/tap/quickai

As a Claude Code plugin (adds the MCP server, ask from a chat)

/plugin marketplace add AlexGladkov/claude
/plugin install quickai@alexgladkov

The plugin registers the quickai MCP server, so you can just say "profile the my-app project" in a chat. It calls the quickai binary — install it first (Homebrew above, or cargo install --git https://github.com/AlexGladkov/quickai) so it's on your PATH.

Prebuilt for Apple Silicon (installs in ~1s); Intel-mac / Linux build from source (needs Rust).

From source

Requires Rust (1.75+). SQLite is bundled — no system deps.

git clone https://github.com/AlexGladkov/quickai.git
cd quickai
cargo build --release
# binary: ./target/release/quickai   (optionally: cp it into ~/.local/bin)

First run

Build the index from ~/.claude/projects/**/*.jsonl:

quickai index

The index lives at ~/.claude/quickai.db (pure derivative of the transcripts — delete and rebuild anytime). Re-run index whenever you want fresh data; it's incremental (only reads the tails of changed files).

2. Example

# overall summary (tokens first, $ as API-equivalent)
quickai stats

# your actual prompts, biggest first
quickai tasks --by tokens --limit 20

# why was a run so long? wall vs idle time
quickai slow

# generate the full HTML report and open it in the browser
quickai report

# report for a single project
quickai report --project my-app

quickai stats output:

quickai — index summary
  projects:   18
  sessions:   97
  tasks:      3182
  subagents:  1988 (type resolved for 981 — 49%)
  turns:      185699
  ── tokens ──
  total:      40709.4M  (input+output+cache)
    fresh (in+out): 135.2M
    input:    31.4M
    output:   103.8M
    cache-read: 39461.8M (96% throughput — context re-reads)
  ≈ API-equivalent: $32042.66  (list-price; subscription is flat — this is volume, not a bill)

3. Screenshots

Report — summary & breakdowns (dark Notion-style theme, project name up top):

Behavior profiling — tool error rates, cache-hit health, spinning, waste, slow tasks (problem numbers are colored at a glance):

Per-task drill-down — click any task to expand its subagents with cost / turns / wall time:

4. What you can generate

CLI commands

The HTML report bundles

Summary cards · by model / project / agent-type · tool profile · latency · parallelism · cache-hit health · spinning · waste · slow tasks with idle % · and every task with a searchable, sortable table and expandable subagent breakdown.

MCP (ask from a chat)

claude mcp add quickai -- /absolute/path/to/target/release/quickai mcp

Then ask things like "profile the my-app project", "what ate the most this week", "break down task X", "why was that run so slow". Most tools accept a project filter, so "profile this project" just works. Tools: quickai_stats, quickai_top, quickai_tasks, quickai_usage, quickai_task, quickai_bench, quickai_cache, quickai_latency, quickai_spin, quickai_tools, quickai_waste, quickai_slow.

Notes

• Money is an API-equivalent (list-price × tokens), not a bill — on a subscription there are no per-token charges. It's a volume/ROI signal.
• No "% of subscription capacity" — Anthropic doesn't publish a fixed token cap (limits are windowed and metered opaquely), so any such % would be made up.
• Claude Code only for now. The core (schema, queries, report) is source-agnostic — only the parse layer is Claude-specific, so other CLIs could be added via adapters.

Architecture & data model: ARCHITECTURE.md.

License

MIT — see LICENSE.