Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Weigh an agent setup and print a fight-card receipt. Use when asked to weigh in, weigh my agent, check my agent's weight or token overhead, or compare before and after a cut. Measures the always-on context bill (installed skills, memory files, MCP tool schemas), assigns a weight class, prints a shareable card, and shows the delta against the last weigh-in. Read-only except for its own append-only log.
| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-05 | ✗→✓ | ▲ Improved | 57% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 361% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 44% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 20% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 35% | 0% |
The scale for your agent. skill-diet finds what to cut; weigh-in tells you what you weigh, before the cut and after it. Step on the scale first, cut, step on it again.
(frontmatter name plus description each), memory files loaded at session start (CLAUDE.md, AGENTS.md, local variants, imports), MCP tool schemas (eager-loaded only; a server you cannot enumerate is unknown, never zero).
the Opus 4.7 tokenizer (Opus 4.7+, Sonnet 5, Fable 5), characters / 3.5 to 4 on older Claude and GPT-style tokenizers. Label estimates.
.cutweight/weigh-ins.md if it exists. Computethe delta.
Cutweight's own scale, by always-on tokens per message:
| class | always-on | |---|---| | flyweight | under 1,000 | | lightweight | under 3,000 | | middleweight | under 8,000 | | heavyweight | under 20,000 | | superheavyweight | 20,000 and up |
textcutweight weigh-in: 2026-07-10 always-on: 7,412 tokens on every message skills: 3,509 (27 installed) memory files: 3,903 mcp schemas: ~0 (deferred loading) class: middleweight last weigh-in: 9,180 (-19%, was heavyweight) next cut: the 25-skill plugin billing 2,462 for 4 skills used
Keep the card under 12 lines. It is made to be posted.
memory) as the next cut. Bill it on the card, flag it to its owner. Its lines are the agent's retrieval keys; hand-trimming them breaks recall.
server, or plugin names. Generalize anything internal before it leaves the machine.
.cutweight/ out of version control; the log stays on the machine.Other measured skills in the registry, with their headline benchmark lift.