Leanfare

Leanfare

Leanfare makes your Claude plan last longer. Claude keeps the hard parts of a task and hands the simple ones, like reading files, searching and drafting, to free or cheaper models. You keep working in Claude as before.

Leanfare is early and changes every week. If something breaks, report a bug; Settings, Report a problem, fills in your version. Questions and ideas go to Discussions.

How it works

1. You ask Claude for something, in Claude Desktop, Cowork or Claude Code.
2. Claude asks Leanfare for a plan. Leanfare splits the task into parts and picks a model for each: free or cheap models for reading and drafting, Claude's own models for the hard parts.
3. Leanfare sends its parts to those models and hands the answers back. Claude does the rest, checks every answer, and answers you.
4. Leanfare's report shows what each prompt used and what it saved.

Get started

On macOS or Linux, paste this in Terminal:

curl -fsSL https://raw.githubusercontent.com/pouriaetab/leanfare/main/install.sh | sh

On Windows, paste this in PowerShell:

powershell -ExecutionPolicy Bypass -c "irm https://raw.githubusercontent.com/pouriaetab/leanfare/main/install.ps1 | iex"

Then type leanfare in the same window. In the page that opens, press Turn on Leanfare, then restart Claude. That's all.

Leanfare needs at least one model to hand work to. It uses the API keys and Ollama models already on your computer, and Settings connects free ones with one press each: ChatGPT (a free account works), Gemini, Groq, OpenRouter, Mistral and Cloudflare.

Other ways in: the Claude Code plugin (/plugin marketplace add pouriaetab/leanfare, then /plugin install leanfare@leanfare), or a one click .mcpb file for Claude Desktop from the latest release.

What you see

The report. Every prompt with its main model, Claude's helpers and the outside models, and what was saved. By prompt or by model, over any days.

Claude's own models. Leanfare picks the Claude model for each helper Claude starts, from llmledger's ratings and prices. As the ratings stand: Fable 5.1 for the hardest parts, Opus 5.5 for high effort, Sonnet 5.5 for medium and Haiku 5.5 for light code. Cowork can ask you to approve each helper, so you choose how many a prompt starts: one per part, one per prompt, or none.

Three settings

1. Level, from 1 (Off) to 5 (All it can): how much Claude hands to Leanfare first.
2. Helpers: one per part, one per prompt (the default), or none.
3. Where work may go: only this computer, or the providers you ticked.

In Cowork and Claude Desktop, a row just above the message box sets the level (or Off), the speed, the quality and the models for the chat you are in. Leanfare picks the rest for you. The guide explains every part.

What's inside

Leanfare is an MCP server with a small web app, both on your computer. Claude connects to it over stdio, like any MCP server. It is TypeScript on Bun, with no runtime dependencies.

File	What it is
src/mcp.ts	The MCP server and its tools
src/cli.ts	The leanfare command
src/server.ts	The app and its API, at http://127.0.0.1:4747
src/planner.ts	Plans: the parts, a model for each, the runs
src/delegate.ts	Hand offs to outside models, each with a receipt
src/claudeladder.ts	Claude's own models, in order
ui/	The app's screens

The MCP tools: plan, plan_run, plan_status, plan_stop, delegate, ask, grade, level, report, spend, models, estimate, trial, local_model, live and helper.

To build and run it from source:

git clone https://github.com/pouriaetab/leanfare.git
cd leanfare
bun install
bun test             # against fake providers; nothing leaves your computer
bun src/cli.ts       # the app
bun src/cli.ts mcp   # the MCP server, over stdio

In any MCP host

Leanfare is a standard MCP server over stdio: the command is leanfare mcp, with no environment variables. When the Leanfare app is not running, the server starts it.

{
  "mcpServers": {
    "leanfare": {
      "command": "/Users/you/.leanfare/bin/leanfare",
      "args": ["mcp"]
    }
  }
}

The install line puts the program in ~/.leanfare/bin/leanfare on macOS and Linux, and in %LOCALAPPDATA%\Leanfare\bin\leanfare.exe on Windows. Use its full path: an app opened from the Dock or the Start menu may not see your shell's PATH. VS Code takes the same command and args in .vscode/mcp.json, under servers with "type": "stdio".

Hosts. Leanfare is built and tested for Claude Desktop, Cowork and Claude Code. There Turn on Leanfare sets it up in one press, and the level, the hooks and the per-prompt report work. Other MCP hosts that run local stdio servers, such as VS Code with GitHub Copilot's agent mode, can run the same tools with the configuration above: the host's model plans with Leanfare, hands it reading and drafting, grades the answers, and the report counts those hand offs. The one-press setup, the level hooks and the main model's own token counts are for Claude's apps only, and other hosts are not tested.

In the MCP Registry. io.github.pouriaetab/leanfare lists an MCPB bundle (registry type mcpb, transport stdio) for each computer Claude Desktop runs on: macOS on Apple silicon, macOS on Intel, and Windows. Each comes from the release with its SHA-256, holds the compiled program, and starts leanfare mcp. For Linux, or a host that does not take MCPB bundles, use the install line and the configuration above.

Privacy

Leanfare reads usage numbers, never your words, unless you choose in Settings to keep a hand off's words on your computer to read later. Your keys stay on your computer, and requests go from your computer straight to the AI company you use. No telemetry. The guide has the details.

Update, stop or take it out

1. Update: run the install line again. Leanfare says when a new version is out.
2. Stop: type leanfare stop.
3. Off for a while: the switch at the top of the app.
4. Take it out of an app: Settings, In your apps, then Remove next to the app.

Data and credits

Prices come from llmledger (MIT), built from LiteLLM's public price list (MIT). Quality ratings come from the Arena leaderboard (CC BY 4.0). The guide lists the rest. Leanfare is not affiliated with any AI provider.

Licence

MIT