claude-eco — eco mode for Claude Code

Claude Code spends your money on work you never asked for: re-reading files it just edited, pasting back diffs it already applied, writing three-paragraph reports for two-line fixes. I measured every token of it, deleted it, and published all 267 raw runs — including the ones where I lose.

/eco: −35% to −57% output tokens on the same review task across Opus 5, Sonnet 5, Opus 4.8 and Fable 5 (n=5 per arm, measured the same day), up to −73% on long agentic tasks at high effort — with both planted bugs found in every single run. Haiku is the documented exception. /eco-max: up to −75% by dialing reasoning effort down — opt-in, labeled. Since 1.2.0 the same rules also ship as an output style (zero invocation cost), plus opt-in hooks that enforce the expensive rules mechanically instead of asking the model nicely.

Quickstart · Results · What you get · Benchmarks · FAQ

Quickstart

/plugin marketplace add sup3x/claude-code-eco
/plugin install claude-eco@claude-eco
/eco

If it answers exactly Eco mode active., it is on for the rest of the session. That is the whole setup.

The waste it removes isn't occasional — it repeats on every turn of every session, which is why usage limits evaporate and API bills surprise you.

What you get

Component	What it does	Cost when idle
/eco	Frugality rules with a non-negotiable quality floor. Full reasoning depth. One invocation covers the session.	~80 tok
/eco-max	The same rules plus a low reasoning-effort override for routine chores. Effort is the single biggest token lever on modern Claude models, and of every terse-mode tool we surveyed this is the only one that touches it.	~80 tok
Eco output style	The identical rule set delivered in the system prompt instead of a skill: no invocation turn, active from the first message, unaffected by compaction. Measured at −36% on the review task, against −57% for invoking the skill — an always-on floor, not a replacement. /output-style Eco.	0 until enabled
/eco-audit	Reads your real settings.json, CLAUDE.md files and MCP configuration and prints what each one costs you, worst first, with the exact settings edit to fix it. Read-only; it never writes without your confirmation.	~110 tok
/eco-report	Per-session token accounting from your local transcripts: output, thinking, cache read/write, out-per-turn, cache health. Observational, and labeled as such.	~80 tok
eco-scout agent	A bounded read-only sweep agent on Haiku at low effort — the delegation the rules recommend, made concrete.	~70 tok
Enforcement hooks	Opt-in PreToolUse/PostToolUse hooks that cap runaway command output, refuse full-file Write over a large existing file, window huge Reads, and bound Grep. Prose the model may ignore becomes mechanism it cannot.	0 — hooks are harness-side and never enter the context

Everything above is one plugin, and claude plugin details claude-eco prints these numbers on your own install: ~419 tokens of component descriptions at startup, bodies loaded only when invoked. The skills work in any language — the rules are English, the replies follow yours.

Measured results