espresso

---

The Problem

Every AI coding agent produces verbose output by default. "Sure! I'd be happy to help. Let me walk you through this step by step..." That's tokens you pay for, time you waste reading, and context window you burn.

Fixing it requires configuring multiple tools, writing rules, setting up hooks. Most developers don't bother.

What Espresso Does

Installs once. Detects what you already have. Adds only what's missing.

What gets configured	Savings	How
Output rules	40-60%	Enforces 120 char lines, forbidden openers/closers, result-first, no filler
Global rules	context savings	Creates ~/.claude/rules/ — Exa search, clean git, GitNexus, project rules
GitNexus	fewer file reads	Configures MCP server + auto-reindex hook (if GitNexus binary installed)
Caveman ultra	~75%	Sets compressed conversation mode (if Caveman plugin installed)
RTK hook	60-90% CLI	Adds CLI output compression hook (if RTK binary installed)
Ponytail	less code written	Installs the real ponytail plugin: YAGNI ladder, stdlib/native first, no speculative abstraction
Model ladder	cost + quota	Spawns every subagent one tier below the session model (Opus runs Sonnet workers, etc.) via a PreToolUse hook

Detection-first: Espresso checks what's already configured and skips it. Never overwrites your existing rules or config. Never installs duplicates.

---

Before / After

Without Espresso:
  Sure! I'd be happy to help you with that. The issue you're
  experiencing is likely caused by a misconfiguration in your
  authentication middleware. Let me explain what's happening and
  walk you through the solution step by step...

With Espresso:
  Bug in auth middleware. Token expiry check uses `<` not `<=`. Fix:

Same information. 70-85% fewer tokens with the full stack.

Three axes, not one

Caveman compresses how Claude talks. Ponytail compresses how much Claude writes. Verbose prose is cheap next to the real waste: an agent that builds a 120-line cache class, adds a dependency, scaffolds "for later", then iterates on its own bloat. Ponytail stops it at the first solution that works, before the wrong code is ever written. Espresso installs both so the axes stack.

The third axis is cost per token, not token count. The model ladder spawns every subagent one tier below the session model: an Opus session dispatches Sonnet workers, a Fable session dispatches Opus workers. Discovery, mechanical checks, and verbose-output tasks do not need top-tier reasoning, so the expensive model keeps judgement and synthesis while cheaper workers do the rest. Each tier down roughly halves cost per MTok and preserves the capped quota of Opus and Fable. Fable 5 is the priciest tier and drains fastest, so a Fable session dispatching Opus workers is where the ladder saves most. Paired with a delegation rule that says when to spawn, tasks offload the expensive tier automatically, at equal quality on bounded work.

---

Install

Claude Code (in the Claude Code prompt)

3 commands. Type them inside Claude Code, not in a regular terminal.

/plugin marketplace add mirkobozzetto/espresso
/plugin install espresso@espresso
/reload-plugins

Or from a regular terminal:

claude plugin marketplace add mirkobozzetto/espresso
claude plugin install espresso@espresso

Then restart Claude Code. First session auto-configures the full stack.

Cursor / Windsurf / Copilot / Codex / Others

These agents don't have plugin hooks. One command in your project root:

curl -sL https://raw.githubusercontent.com/mirkobozzetto/espresso/main/AGENTS.md > AGENTS.md

Codex, Cursor, Windsurf, Copilot, Amp, and Devin read AGENTS.md natively. You get the output rules (40-60% savings) but not the full stack auto-install.

---

Exa MCP — Web Search for Your Agent

Espresso rules enforce Exa as the only web search tool. Exa is a free hosted MCP — no API key required.

Agent	Setup
Claude Code	claude mcp add --transport http exa https://mcp.exa.ai/mcp
Codex	Add exa MCP in .codex/config.toml with URL https://mcp.exa.ai/mcp
Cursor	Add to ~/.cursor/mcp.json: {"mcpServers": {"exa": {"url": "https://mcp.exa.ai/mcp"}}}
VS Code	Add to .vscode/mcp.json: {"servers": {"exa": {"type": "http", "url": "https://mcp.exa.ai/mcp"}}}
Claude Desktop	Settings → Connectors → search "Exa" → click +