Skip to content
agentsetups
// insights · data story

The lean setup movement.

Read enough real setups and a pattern shows up: people get better results by taking things out. Here's what our 50 cards say. Every number below is a count of our cards or a figure quoted from a linked source, not a survey of everyone.

20/50

cards tagged lean / pruning

38/50

cards that cut, replaced or disabled something

12/25

of the cards that name a model include an older one

01 · the receipts

Numbers people published about their own bloat.

90 of 103

installed capabilities never fired in 30 days

pycomet · source →

2 of 22

MCP servers used daily

Flavio Copes · source →

80%+

of a harness system prompt cut for newer models

Anthropic (reported by onthewire.ai) · source →

11,370 → 1,000

tokens: his Pi cold start vs Pi's floor, gap mostly heavy tools

Kevin Lewis · source →

6.7% → 68.3%

Grok Code Fast 1 after swapping only the edit tool

Can Bölük · source →

23k

tokens gone by swapping the GitHub MCP for gh

Peter Steinberger · source →

02 · what gets cut

MCP servers top the cut list.

Number of our cards that removed, replaced, disabled or flagged each kind of thing (grouped by keyword).

03 · five moves

How lean setups get lean.

Our grouping of what the cards do. Each move links the setups that show it.

  1. 1

    Measure, then delete

    Count what actually fires before you prune.

    pycomet →thousandflowers (GitHub) →Flavio Copes →
  2. 2

    CLIs over MCPs

    If a CLI exists, the model already knows it, and it costs nothing until called.

    Peter Steinberger →Armin Ronacher →Kevin Lewis →
  3. 3

    Off by default

    Load tools per session or per path, not globally.

    Kevin Lewis →Luca Cavallin →Thomas Tartrau →
  4. 4

    Tiny always-on rules

    Short top-level instructions; detail in scoped rules and skills.

    Thomas Tartrau →Philip Mathew Hern →llmx.tech →
  5. 5

    Fewer tools, better format

    A minimal toolset, and fixing tools before upgrading models.

    Mario Zechner →Can Bölük →DeepSeek →
04 · harnesses

Where these setups run

Cards naming each harness. Many name more than one.

05 · current models

Current models named

Cards naming a model on our current list. Most cards don't name one.

  • Claude Opus 5.53
  • Claude Sonnet 5.52
  • GPT-6.1 Sol2
  • DeepSeek V4.1 Flash2
  • GPT-6 Astra on Codex1
  • GPT-6 Luna1
  • GPT-6.1 Sol, high reasoning1
  • GLM 5.3 / GLM 5.3 Flash1
  • GPT-6 Astra1
  • Grok 4.71
  • GLM-5.31
  • DeepSeek V4-Pro-08131

Is your setup lean?

Build a starter config from the lean setups above.

▸ Build a lean rig