← Back to Discover

Entroly

Local-first tool that sits between an AI coding agent and its model provider, selecting repository context under a token budget, recording what was left out and checking answers against the evidence.

MCP serversCommand-line tools
Visit website ↗

What it does

Entroly sits between an AI coding agent and its model provider. It ranks the files in a repository, selects the context that fits a token budget, and sends the rest as signatures or references that can be expanded on demand. Each selection can produce a Context Receipt listing what was used, what was left out, why, and which risks remain. It can wrap an agent, run as a local HTTP proxy that you point a base URL environment variable at, act as an MCP server, or be imported as a Python library. The README lists 38 supported agent targets, including Claude Code, Codex CLI, Aider, Gemini CLI, Cursor, Windsurf, VS Code and Zed. Other parts include keeping the prompt prefix stable so provider caches keep hitting, optional routing of easy repeated tasks to cheaper models, and a verifier called WITNESS that checks an answer against the evidence it was given. The README reports token savings and accuracy tables, for example an average 87% token reduction across its own workloads, and says they can be reproduced from committed scripts; they are the project's own figures. It says code is indexed and selected on the device, with no outbound analytics by default, under the Apache License 2.0. It installs with pip, npm, Homebrew or Docker, and it passes content through unchanged for tiny repositories or short prompts that already fit the budget.

Capabilities

  • Selects repository context under a token budget while compressed fragments stay retrievable
  • Context Receipts show what was selected, what was omitted and why
  • Wrap an agent, run a local proxy, plug in as an MCP server or use as a library
  • WITNESS verifier checks answers against the supplied evidence
  • Keeps the prompt prefix stable so provider caches can keep hitting
  • Optional routing of easy tasks to cheaper models

When to use it

  • Your agents work on large repositories and only see a few files at a time
  • You want to reduce the token cost of AI coding
  • You need an inspectable record of what context an answer was based on
  • You want answers checked against evidence before you trust them