📚 New: the Encyclopedia of Token Efficiency — every technique to cut your agent's token bill, explained.
The world leader in token efficiency

Token efficiency
is all we do.
Nobody does it better.

SuperLeanAI is a token efficiency company — full stop. AI labs have every economic incentive to maximize your token usage; nobody was pushing back until we did. We debloat, trim, and minify every request your agent sends — built-in tools, MCP servers, Skills, all of it — so you spend as few tokens as possible. We only make money when you save tokens.

Also known as prompt compression, context compression, or prompt debloating — whatever you call it, we do it for every request your coding agent sends.

Same task, same agent — side by side, with and without SuperLean.

20%
reduction in token bills for $1 a month
0
code changes needed in your agent
30+
tools debloated per agent session
4
coding agents supported (Claude Code, etc)

Every request bleeds tokens —
token efficiency fixes that

Agents like Claude Code send lots of information (10+ tool definitions, plus whatever your MCP servers and Skills add) on every LLM call. The labs have no incentive to make them short — token efficiency is our entire business.

✗ BEFORE — original Bash tool 4,839 chars · ~1,210 tokens
{
  "name": "Bash",
  "description": "Executes a given bash command with optional
  timeout. Working directory persists between
  commands; shell state (everything else) does not.
  The shell environment is initialized from the
  user's profile (bash or zsh).

  IMPORTANT: This tool is for terminal operations
  like git, npm, docker, etc. DO NOT use it for
  file operations (reading, writing, editing,
  searching, finding files) — use the specialized
  tools for this instead.

  Before executing the command, please follow
  these steps:

  1. Directory Verification:
     - If the command will create new directories
       or files, first use `ls` to verify the
       parent directory exists ...

  2. Command Execution:
     - Always quote file paths that contain spaces
       with double quotes ...
  [... 4,500 more characters of instructions ...]",
  "input_schema": {
    "properties": {
      "command": { ... },
      "description": {
        "description":
          "Clear, concise description of what this
          command does in active voice. Never use
          words like 'complex' or 'risk'...
          [200 more chars of meta-instructions]"
      },
      "timeout": { ... },
      "run_in_background": { ... }
    }
  }
}
✓ AFTER — debloated by SuperLeanAI ~380 chars · ~95 tokens · saved 92% per tool call
{
  "name": "Bash",
  "description":
    "Executes bash commands with
    optional timeout.",
  "input_schema": {
    "properties": {
      "command": { ... },
      "description": {
        "description":
          "The bash command to execute."
      },
      "timeout": { ... },
    }
  }
}






~1,210
tokens (before)
~95
tokens (after)
×
30+
tool calls per session
=
33,000+
tokens saved per session
×
$15
/ M tokens (Opus)
=
$0.50
saved per session
×
100
sessions / day
=
$50
savings per day

Three things happen on every request

SuperLeanAI is a transparent HTTP proxy. Point your agent's BASE_URL to api.superleanai.com (e.g. via ANTHROPIC_BASE_URL) and you're done.

✂️
STEP 01

Tool, MCP & Skills debloat

Verbose, multi-paragraph tool descriptions — from built-in tools, MCP servers, and Skills alike — are summarized to a single sentence. The model still understands the tool — it just costs 92% less to say so.

🚫
STEP 02

Tool & MCP filtering

Block tools, MCP servers, or Skills your agent doesn't really need (e.g. NotebookEdit, Grep, an unused MCP server). Or use agressive bash_only mode: strip everything down to just Bash, Read, and Edit — the minimum set for most coding tasks.

🔀
STEP 03

On the fly model switching & routing

Override the model on the fly — swap Opus for Sonnet during bulk tasks, or route different profiles to cheaper models without touching agent code.

Everything your agent bill needs

Built as a production-grade proxy with JWT auth, per-session stats, and live savings.

📊

Per-session savings tracking

Every session gets live stats: original tool bytes, debloated bytes, total traffic, and percentage saved. Know exactly what you're not paying for.

🧩

MCP & Skills debloat

MCP servers and Skills are often the biggest source of bloat — every connected MCP tool schema and every loaded Skill's instructions ship on every call. Debloated the same way as built-in tools, transparently.

🔌

Multi-backend routing

Route to OpenRouter, Google Gemini (via googleapis), Groq, or Together AI. Configure named profiles — groq, googleapis_minimal — with different strategies per use case.

🗂️

Traffic logging

Every request is saved to disk as JSON. Replay, audit, or analyze your agent's traffic. Tool calls from the response are also captured for future debloating.

🔧

Drop-in compatible

Speaks the Anthropic /v1/messages API. Works with Claude Code, custom agents, any client targeting the Anthropic SDK. No code changes required.

Running in one command

Stay logged into Claude Code or Codex and start saving with npx superlean-claude or npx superlean-codex — no env vars, no config. Pick your agent below.

Prompt compression, context compression, whatever you call it

Straight answers to how this actually works.

What is prompt compression?

Prompt compression — also called context compression or prompt debloating — is shrinking the text an LLM API call sends (tool definitions, MCP server schemas, Skills instructions, system prompts, history) while keeping its meaning, so each request costs fewer tokens.

Does it compress MCP servers and Skills, or just built-in tools?

All of it. MCP servers and Skills often add more tokens per request than an agent's built-in tools, since every MCP tool schema and every loaded Skill's instructions go out on every call. SuperLeanAI debloats built-in tools, MCP servers, and Skills the same way, transparently.

How is SuperLeanAI different from a generic context compression tool?

SuperLeanAI is a transparent HTTP proxy purpose-built for AI coding agents. It debloats verbose tool definitions, MCP server schemas, and Skills instructions, filters what's unused, and trims context on every request with zero code changes — point your agent's base URL at the proxy and it works.

Does prompt compression change model behavior?

No. Verbose descriptions are summarized to their essential meaning, not truncated, so the model still understands every tool, MCP server, and Skill — it just costs far fewer tokens to say so.

Which agents does it work with?

Claude Code, Codex CLI, GitHub Copilot CLI, and Cline today — anything that speaks the Anthropic or OpenAI message APIs works out of the box.

We win when you save.

Our incentives are completely aligned with yours. We take a small cut of the savings — if you don't save, you don't pay. Crystal clear, no subscription traps.

Get started →