Deadweight: What your Claude Code setup costs on every request A tool called deadweight, installed via `cargo install deadweight` or a prebuilt binary for macOS, Linux or Windows, audits Claude Code session history to measure the token overhead that plugins, skills, MCP servers and hooks add to every request before the user types anything. In the example run, each request carried 61.2k tokens (median of recent sessions 63.8k), of which ~19.4k came from the user's setup and ~41.8k from Claude Code itself, with deferred tool names (412) at ~7.1k tokens and skill listings (214) at ~6.2k the largest contributors. The tool reports that Claude Code caps the skill listing at roughly 30,000 characters, so 151 of 214 skills were listed by name only without descriptions, and it offers a `deadweight fix` command that disables 6 unused plugins and hides 11 unused skills to save ~6.8k tokens per request. Find the dead weight in your Claude Code setup. Every plugin, skill, MCP server and hook you install adds text to every single request Claude Code sends, before you type a word. deadweight measures that overhead from your own session history, tells you exactly where it comes from, shows what you never actually use, and removes it with one undoable command. An example run. Yours will look different, and that's the point. Same output as text bash $ deadweight Before you type a word, every request carries 61.2k tokens measured from your latest session; median of recent sessions 63.8k ~19.4k from your setup · ~41.8k from Claude Code itself system prompt + 18 tools WHAT YOUR SETUP ADDS Deferred tool names 412 ~7.1k ████████████████████████ Skill listings 214 ~6.2k █████████████████████ Hook injections 3 ~3.4k ███████████▌ MCP server instructions 4 ~2.1k ███████▏ Agent listings 12 ~600 ██ BIGGEST SOURCES name type tokens used contents github mcp server ~3.9k 212× 91 tools kitchen-sink plugin ~3.6k never 64 tools · 22 skills workflow-kit plugin ~2.6k 14× hooks 2.2k · 11 skills playwright mcp server ~1.4k never 33 tools devops-pack plugin ~1.3k 3× hooks 1.2k · 6 skills marketing-suite plugin ~1.1k never 18 tools · 14 skills linear mcp server ~810 37× 41 tools PROBLEMS ✗ 151 of 214 skills are listed without a description: the skill listing is full ~30.0k characters . Claude sees only their names, so it can't tell when to use them. Including 4 you actually use: pdf-tools:extract, review-pr, workflow-kit:plan, +1 more Make room: hide what you don't use deadweight fix , or set "skillListingBudgetFraction": 0.02 in settings.json default 0.01; the extra room costs up to ~8.3k tokens per request . ✗ 2 MCP servers failed to connect at startup 9 MCP servers never signed in, still listing 18 auth-only tools 3 skill names are listed more than once 198 of 214 listed skills were not used in 30 days ~5.1k per request Run deadweight fix to disable 6 unused plugins and hide 11 unused skills and save ~6.8k tokens per request. UNUSED, BUT YOURS TO REMOVE mcp server playwright ~1.4k claude mcp remove "playwright" LOADED IN EVERY PROJECT, USED IN ONE OR TWO mcp server linear ~810 37× in /home/you/code/app of 9 projects in .claude.json, move "linear" from the top-level mcpServers into projects "/home/you/code/app" mcpServers - Your usage limit drains faster. The overhead is resent on every turn of every session. - Your context window fills sooner , so sessions compact earlier and Claude loses track of your actual work. - Claude gets worse at picking the right skill. This one surprised us: Claude Code caps the skill listing at roughly 30,000 characters. Once you go past it, skills are listed by name only, without their descriptions . With a few hundred skills installed, most of them end up as bare names, and Claude has no way to tell when to use them. Unused skills crowd out the ones you rely on. None of this shows up anywhere in the UI. Plugins get installed one at a time and the total is never visible. cargo install deadweight Or grab a prebuilt binary for macOS, Linux or Windows from Releases https://github.com/msertdev/deadweight/releases . It's a single binary with no runtime dependencies. deadweight audit: what every request carries and where it comes from deadweight --all list every source, not just the biggest deadweight --days 60 count anything used in the last 60 days as "used" deadweight --json machine-readable output deadweight fix --dry-run show what would change deadweight fix disable unused plugins, hide unused skills asks first deadweight undo put everything back deadweight card save your totals as a PNG to share no names on it fix options: | flag | effect | |---|---| | --keep NAME | never touch this plugin or skill, or connectors repeatable | | --park-skills | move unused skills into the backup folder instead of hiding them | | --include-hooked | also disable unused plugins that run hooks | | --yes | don't ask for confirmation | | --force | run even with fewer than 5 sessions of history | Restart Claude Code after fix or undo to load the new setup. deadweight card writes deadweight-card.png : your total, how much of it is your setup, unused skills, and what fix would save. It shows totals only, never a plugin, skill or server name. --output card.svg gives you the SVG instead. Post yours. I'd like to see the biggest one. deadweight ships as a Claude Code plugin too, so you can ask Claude to audit its own setup: /plugin marketplace add msertdev/deadweight /plugin install deadweight@deadweight /deadweight The plugin needs the deadweight binary on your PATH see Install install . It adds a single short line to the skill listing, about 18 tokens. Claude Code keeps a transcript of every session in ~/.claude/projects/ . Alongside the conversation it records what was loaded at startup: the skill listing, the deferred tool names, MCP server instructions, the agent listing, hook output, and the token usage the API billed for each request. deadweight reads those records: 1. Measure. The total comes from the API usage of the first request of your latest session input + cache write + cache read tokens . This number is measured, not estimated. 2. Attribute. Each skill line, tool name, instruction block, agent and hook injection is traced to the plugin, skill, MCP server or connector that added it. Hook commands are matched against each plugin's hooks.json . Sizes are converted to tokens from text length; whatever your setup doesn't explain is Claude Code's own system prompt and tools. 3. Check usage. Every transcript from the last 30 days including subagents is scanned for Skill calls, MCP tool calls, agent launches and slash commands. 4. Diagnose. Failed and never-authenticated MCP servers, duplicate skill names, skills that lost their description, and single skills with very large descriptions. 5. Scope. A user-scope MCP server loads in every project. If its tools were only ever called from one or two projects, deadweight says which ones, so you can move it to those projects and stop paying for it everywhere else. A project's .mcp.json works too, if the team should share it. On the author's machine, 1.4 GB of transcripts 41 sessions, 400+ subagent runs are scanned in about a second, in parallel. - Read-only by default. Only fix and undo write anything, and both ask first. - No secrets are read. From .claude.json deadweight takes only MCP server names , never commands, environment variables or headers. A test checks that secret values never reach the output. - Nothing is deleted. Plugins are disabled in settings.json enabledPlugins: false , not uninstalled. Unused user skills are hidden from the model with skillOverrides "user-invocable-only" , so /name still runs them; --park-skills moves them into the backup folder instead. If none of your claude.ai connectors was used, disableClaudeAiConnectors switches them off in Claude Code they stay connected on claude.ai . - Everything is backed up. Each fix writes a manifest and a copy of settings.json to ~/.claude/deadweight/backups/