juyterman1000/entroly

Compress tool outputs, logs, files, conversations, and RAG context before they reach the model. On measured workloads, Entroly reduces unnecessary tokens by up to 90% while preserving answer-critical evidence through budget-aware selection, content-addressed recovery, and auditable Context Receipt. Works with Claude, Chatgpt/Codex,Openclaw.

What it solves

Entroly is a "Context OS" designed to manage the information sent to AI agents, preventing them from crashing or becoming too expensive due to context window overflow. It solves the problem of "runaway sessions" where autonomous loops append logs until the provider rejects the request, and ensures that when context is compressed to save tokens, the most critical evidence is preserved and remains recoverable.

How it works

Entroly acts as a local observability and decision layer that sits between the AI agent and the LLM provider. It uses a combination of BM25, entropy scoring, SimHash deduplication, and dependency graphs to rank and select the highest-value evidence under a specific token budget.

Key mechanisms include:

  • Recoverable Compression: It keeps original versions of omitted text locally via content-addressed handles, allowing the agent to retrieve exact chunks if needed.
  • Proxy Guard: When used as a proxy, it monitors outbound requests and compacts repetitive tool output in a "live zone" while preserving structural metadata and errors.
  • WITNESS: A local verification system that checks model answers against supplied evidence without requiring additional model calls.
  • Stable Prefixing: It maintains stable system and historical bytes to protect native provider cache boundaries, reducing redundant processing.

Who it’s for

AI agent developers and enterprise teams who use tools like Claude Code, Cursor, Aider, or custom MCP clients and need to reduce API costs, prevent session crashes, and maintain an auditable trail of what the model actually read.

Highlights

  • Local-First Architecture: Runs locally with Python (optional Rust acceleration) or Node/WASM.
  • Comprehensive Tooling: Provides a CLI, SDK, MCP server, and Proxy for flexible integration.
  • Evidence Receipts: Generates detailed reports on what was included, omitted, and the potential risks associated with each selection.
  • Multi-Provider Support: Compatible with OpenAI, Anthropic, Gemini, and local providers like Ollama/LM Studio.