# juyterman1000/entroly [Health: Active]

**Category:** 🧠 Knowledge & Memory  
**Repository:** https://github.com/juyterman1000/entroly  
**GitHub Stars:** 443  
**Views:** 3  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/juyterman1000-entroly

## Description
Auditable context control plane and MCP server for AI coding agents. Compresses context 70–95% (BM25 + entropy + dep-graph knapsack), stabilizes prompt prefixes for provider cache discounts, routes easy tasks to cheaper models (RAVS Bayesian router), and verifies answers locally with WITNESS hallucination guard (0.844 AUROC, $0, 3 ms). MemoryOS adds local budget-aware working/episodic/semantic memory with decay, safety scanning, and durable persistence. 38 agent integrations (Cursor, Claude Code, Codex, Aider, and more). Ships as MCP server (entroly serve), HTTP proxy, or Python/Rust library. Apache-2.0, local-first. pip install entroly

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "entroly": {
    "command": "npx",
    "args": ["-y","entroly@latest"]
  }
}
```

## Documentation

## What juyterman1000/entroly MCP server does

juyterman1000/entroly MCP server sits between an AI coding workflow and its context sources. It selects evidence that fits an explicit token budget, compresses the selected context, and keeps omitted material recoverable. The project describes this as a local-first Context Assurance layer rather than a replacement model or agent architecture.

The main use case is reducing unnecessary provider-bound input while retaining an audit trail. Entroly records what was included, what was excluded, and the reasons behind the selection in auditable receipts. Content-addressed recovery is intended to preserve original source spans exactly when omitted evidence needs to be retrieved later.

The repository also describes optional components for model routing, local memory, and answer checking. RAVS routes easier tasks toward less expensive models, while MemoryOS provides budget-aware working, episodic, and semantic memory with decay, safety scanning, and durable persistence. WITNESS is described as a local hallucination guard that verifies answers without an API cost.

## How it works

juyterman1000/entroly MCP server uses several signals to choose context, including BM25 relevance, entropy, and dependency-graph relationships. These signals are combined under a knapsack-style token budget so the resulting prompt favors higher-value evidence. Compression is designed to be recoverable rather than destructive: source material set aside by the selection process remains available for later retrieval.

The system can operate at different integration boundaries. The README lists MCP, proxy, plugin, wrapper, and SDK paths, including direct paths for OpenAI-compatible and Anthropic-compatible SDKs, Vercel AI SDK middleware, LangChain, LiteLLM, OpenClaw, and OpenCode. Provider-bound savings only apply when traffic actually passes through an Entroly-controlled route.

Tool-schema reduction is opt-in through the `X-Entroly-Active-Tools` request header. A caller can provide a comma-separated active tool set, such as `search_files,read_file`. Forced tool choices and unnamed provider tools remain available, and an invalid or non-matching set leaves the request unchanged.

## Setup and configuration

Install the Python distribution with `pip install entroly`, then start the MCP entry point with `entroly serve`. The project is also published on npm, but the supplied material does not specify an npm server command or client configuration block.

The installation runs locally and does not require an API key according to the project’s installation guidance. Exact local token and cost measurements can be viewed with `entroly value`, `entroly value --json`, or `entroly dashboard`. These figures are local measurements; the README distinguishes them from provider invoices.

Entroly can also be used as an HTTP proxy or through library interfaces, but the provided material does not give the commands or configuration fields for those modes. The project is Apache-2.0 licensed.

## Tools and capabilities

juyterman1000/entroly MCP server provides the following documented capabilities:

- Selects high-value evidence within explicit token budgets.
- Compresses context while keeping original evidence recoverable.
- Produces receipts describing context-selection decisions.
- Tracks tokens saved, estimated cost avoided, compression savings, and deferred tool-schema tokens.
- Supports local working, episodic, and semantic memory through MemoryOS.
- Offers local answer checking through WITNESS and model routing through RAVS.
- Integrates through MCP, proxy, plugin, wrapper, and SDK paths.

The README names Cursor, Claude Code, Codex, OpenClaw, GitHub Copilot, Aider, and other agent or application integrations. A listed integration does not by itself mean every hosted inference path is intercepted; the request must use a supported Entroly-controlled route.

## Limitations and notes

Compression savings depend on the workload and on whether requests traverse Entroly. The project’s live metrics are installation-specific, and estimated cost avoidance is a model based on configured pricing rather than a provider billing record. Optional community telemetry is described as privacy-limited and separately opted in; it is not an exact global usage counter.

The README does not identify individual MCP tool names or provide a complete client-by-client setup matrix in the supplied material. It also does not establish that every named integration supports provider-bound savings. Developers evaluating the MCP mode should verify the expected client configuration and the behavior of their particular request route before relying on compression or cost measurements.

_Full upstream README: https://allmcps.com/mcp/juyterman1000-entroly/readme_