claude-trim

Find out what Claude Code loads before you type a single word.

claude-trim is a static analyzer for Claude Code. It scans your ~/.claude/ directory, counts how many tokens every skill and MCP server costs at startup, flags skills that conflict with each other, and tells you exactly what to remove — and how much money that saves.

demo

Why this exists

Every Claude Code session starts with a hidden tax. Before you write a single prompt, Claude loads the description of every skill you have installed. With a typical plugin setup that's 20–30k tokens per session — paid silently, every time.

Most of those tokens don't help you. Duplicate skills cover the same ground. Skills you installed months ago and forgot about run on every session. And when two skills give Claude contradictory instructions, neither works well.

With 55 skills installed:

Startup cost:  23,751 tokens  (paid every session)
Peak cost:    140,391 tokens  (if every skill body loads)

At Sonnet pricing ($3/M), 5 sessions/day, 22 working days:
→ ~$7.84/month in skill overhead alone

claude-trim makes this visible in 10 seconds, offline, no API key required.

Install

go install (recommended)

go install github.com/d0d012/claude-trim/cmd/claude-trim@latest

Make sure $GOPATH/bin (usually ~/go/bin) is on your $PATH:

export PATH="$HOME/go/bin:$PATH"

Download binary

Grab the binary for your platform from GitHub Releases:

# macOS (Apple Silicon)
curl -L https://github.com/d0d012/claude-trim/releases/latest/download/claude-trim_darwin_arm64.tar.gz | tar xz
mv claude-trim /usr/local/bin/

# macOS (Intel)
curl -L https://github.com/d0d012/claude-trim/releases/latest/download/claude-trim_darwin_amd64.tar.gz | tar xz
mv claude-trim /usr/local/bin/

# Linux (amd64)
curl -L https://github.com/d0d012/claude-trim/releases/latest/download/claude-trim_linux_amd64.tar.gz | tar xz
mv claude-trim /usr/local/bin/

Build from source

git clone https://github.com/d0d012/claude-trim.git
cd claude-trim
go build -o claude-trim ./cmd/claude-trim

Quick start

claude-trim scan

That's it. The interactive TUI opens immediately.

Press Tab to switch between the flamegraph and the conflict table. Use ↑/↓ or j/k to scroll. Press q to quit.

JSON output (for scripting)

claude-trim scan --json

{
  "token_report": {
    "system_base": 20000,
    "total": 23751,
    "peak_total": 140391,
    "skills": [
      {
        "name": "skill-creator",
        "tokens": 59,
        "body_tokens": 7068,
        "pct": 0.25
      }
    ]
  },
  "conflict_report": {
    "conflicts": []
  },
  "suggestions": [
    "Remove 'skill-creator' to save ~7068 tokens when invoked (~$0.03/month if used daily)"
  ]
}

Scan a different directory

claude-trim scan --dir /path/to/other/.claude

Show low-severity conflicts too

claude-trim scan --verbose

What the numbers mean

The flamegraph shows two numbers per skill: startup / on-demand

Column	What it is	When you pay it
startup	Description tokens	Every session, always
on-demand	Full body tokens	Only when the skill is invoked
peak%	Share of worst-case total	Reference — not a daily cost

Claude Code loads only the description of each skill into the system prompt at startup. The full skill body is loaded on-demand when Claude calls the Skill tool. MCP server configs are always loaded.

This means the startup cost is what you pay guaranteed. The on-demand cost is what you pay if a skill actually runs.

Features

Feature	Detail
Token flamegraph	Horizontal bar chart sorted by on-demand cost. Bars scaled relative to largest skill. Color = share of peak (red ≥10%, amber ≥5%, green <5%)
Conflict detection	Jaccard similarity on skill descriptions. High ≥60%, Medium ≥35%, Low ≥15%
False positive filtering	Skills from the same plugin are never flagged as conflicting — they're designed to work together
Accurate counting	Startup cost uses description tokens only. On-demand cost uses full body. MCP configs counted separately
JSON output	`--json` flag for piping into `jq`, dashboards, or CI
Fully offline	Uses `tiktoken-go` (cl100k_base approximation). No API calls, no keys
Fast	Single binary, scans 55+ skills in under a second

How conflict detection works

Extract keywords from each skill's description field (lowercase, stopwords removed)
For each pair of skills from different plugins: compute Jaccard similarity
Flag pairs above the threshold

Jaccard(A, B) = |keywords_A ∩ keywords_B| / |keywords_A ∪ keywords_B|

≥ 0.60 → High    (likely redundant, remove one)
≥ 0.35 → Medium  (significant overlap, worth reviewing)
≥ 0.15 → Low     (minor overlap, shown with --verbose only)

Skills in the same plugin directory are skipped — they're intentionally related.

Note: Token counts are approximations. Claude uses a proprietary tokenizer; claude-trim uses cl100k_base via tiktoken-go. Real counts may differ by a few percent.

What this tool does NOT do

No session history analysis — use ccusage for that
No file modifications — read-only in this version
No network calls — fully offline
No support for Cursor, Codex, or other tools — Claude Code only

Contributing

git clone https://github.com/d0d012/claude-trim.git
cd claude-trim
go test ./...

Tests use fixture files in testdata/ — no real ~/.claude/ directory needed.

Conventional commits please: feat:, fix:, docs:, test:, chore:. Releases are cut automatically on v* tags via goreleaser.

Open an issue before a large PR. Small fixes welcome directly.

claude-trim

claude-trim

Why this exists

Install

go install (recommended)

Download binary

Build from source

Quick start

JSON output (for scripting)

Scan a different directory

Show low-severity conflicts too

What the numbers mean

Features

How conflict detection works

What this tool does NOT do

Contributing

License

Yorumlar (0)