l6e

About l6e

l6e is an MCP server that gives a coding agent a spending budget and, in doing so, changes how the agent behaves. The argument behind it is simple and testable: without a cost signal an agent has no reason to be economical, so it reads files it does not need, searches broadly when it could search narrowly, and keeps going past the point where it already knows enough to act. l6e adds a budget gate to the MCP stack — the agent calls l6e_run_start with a dollar budget, l6e_authorize_call before each tool invocation returns allow or deny along with the remaining budget, and l6e_run_end reports total cost and calls made. The claimed effect is not just cheaper runs but better ones, because the checkpoint forces the agent to ask whether it actually needs the next file read. The project's own evidence is that its entire documentation site was built with frontier models for ninety-nine cents across a few two-dollar session budgets, and it quotes the model it was working with describing the checkpoints as clarifying rather than constraining. Privacy is a stated design constraint: l6e never reads prompts, only token counts and estimates. It installs with pip and works with Cursor, Claude Code, Windsurf and any MCP-compatible client, sitting in front of other MCP servers in the same stack.

Key Features

Tags

More AI Agent Infrastructure tools