M
MCP Directory CURATED DIRECTORY
911 MCP servers indexed
中文
← Back to directory
E

Entroly

Community
Recoverable, auditable context compression for AI agents.
How to install ↓ Compare GitHub source repository ↗
Category
Dev Tools #26 of 438
Stars
★ 480 Very popular
Transport
stdio (local process)
Runtime
Python · Node.js
Credentials
No credential needed
License
Apache-2.0
Last commit
67FMRS · C

Entroly targets AI agent developers who need context governance rather than another chat runtime. It provides local-first selection, compression, recovery, receipts, and verification through Python, Node/npm, and stdio MCP surfaces. Actual interception depends on the integration path, recovery depends on retained local state, and token or quality claims should be validated on the target workload.

Strongest · Documentation 17/20 Weakest · Reliability 10/20

Reliability
10/20
Security and permissions
14/20
Maintenance
13/20
Documentation
17/20
Setup experience
13/20
Why each score
Reliability 10/20
The supplied material shows versioned PyPI/npm stdio manifests, verification commands, explicit error boundaries, and several recovery mechanisms, making the MCP startup path plausible. However, no server source, CI workflow, or committed MCP initialization/tool-list tests are provided, so handshake correctness and tool/implementation consistency cannot be verified. The broad feature surface and many unverified claims add uncertainty, so points are deducted.
Security and permissions 14/20
The material describes local-first operation, no outbound analytics by default, user-configured upstream proxy routing, non-persisted credentials, and project-, scope-, and time-limited MCP grants. No malware, credential theft, or unconfirmed delete/pay/deploy red line is evident. However, source and permission-implementation evidence is absent, while the system handles source code, conversations, logs, and recovery stores; retention, filesystem permissions, network boundaries, and confirmation for dangerous operations are not sufficiently demonstrated, so full marks are not justified.
Maintenance 13/20
The project has an Apache-2.0 license, version 1.0.67, a non-archived repository, an identifiable maintainer, and extensive maintenance-oriented documentation. The supplied material does not show commit history, release cadence, dependency update practice, vulnerability response procedures, or a security contact. Zero open issues does not establish sustained maintenance, so points are deducted.
Documentation 17/20
The README covers installation, integration paths, authentication boundaries, recovery, cost-accounting scope, limitations, benchmark scope, and troubleshooting boundaries. It also carefully distinguishes local estimates from provider-observed results, supporting a high documentation score. Points are deducted because individual MCP tool parameters and return contracts, permission matrices, platform dependencies, and independently reviewable troubleshooting evidence are incomplete; several claims depend on source and documents not supplied here.
Setup experience 13/20
PyPI, npm, uvx, Docker, stdio MCP, proxy, CLI, and Claude Code/Codex/OpenClaw examples are provided, with clear mainstream paths and no real secrets in examples. Points are deducted because the many integration choices and Python/Rust/WASM/proxy modes increase complexity, client behavior has explicit mode boundaries, and no source, CI, or real handshake tests are supplied to verify that installation reliably results in a working connection.

Static review · not runListed 2026-07-30

Read the FMRS scoring method →

Fit and risk

What it can accessReads local filesWrites / deletes local filesUses the network

Best for

  • Large repositories and document collections
  • Multi-turn, log-heavy AI agent sessions
  • Teams requiring evidence traceability, recovery, and audit records
  • Developers who want local processing and control over the model integration path

Not for

  • Small projects or prompts already within budget
  • MCP-only setups that require automatic interception of every model request
  • Tasks that require the full context to remain visible at all times
  • Treating scoped benchmark results as guarantees across all models and production workloads

Required permissions

  • Requires access to configured local project code, documents, logs, or other context data
  • MCP attachments can be scoped by project and grant scope, given an expiry, and revoked
  • Proxy mode requires the user to configure the relevant provider API key; Entroly does not choose the provider or receive its credentials

Risks and side effects

  • Compression can omit relevant evidence; review Context Receipts and verification results
  • Exact recovery depends on retained receipts and the local recovery store; deleting that state removes the recovery path
  • MCP integrations do not automatically intercept every model request; provider savings require routing through the Entroly proxy
  • Unknown or ambiguous structures pass through unchanged, and unsafe overflow may produce HTTP 413
  • README benchmarks are scoped to particular datasets, budgets, models, and versions, not universal guarantees

Setup

Before you start

Runtime：Python · Node.js

Other optional settings (6)
ANTHROPIC_BASE_URL optional Optional variable pointing an app at the local Entroly proxy (http://localhost:9377); proxy mode only.
OPENAI_BASE_URL optional Optional variable pointing OpenAI-compatible clients at the Entroly proxy (http://localhost:9377/v1); proxy mode only.
GOOGLE_GEMINI_BASE_URL optional Optional variable pointing Gemini clients at the Entroly proxy (http://localhost:9377/v1beta); proxy mode only.
ENTROLY_LOCAL_NLI optional Set to 1 to enable the optional local DeBERTa NLI verification without an API call.
ENTROLY_NO_DOCKER optional Set to 1 so `entroly serve` uses the installed Python runtime instead of the default Docker image.
ENTROLY_VERIFIED_DREAMING optional Set to 1 to opt in to the experimental verified dreaming self-improvement loop.

Install the Python package with pip install -U entroly. Register Entroly as a stdio MCP server using the command entroly with no arguments; entroly init is also documented. For Claude Code, Codex, or OpenClaw, use entroly attach create --client <client> --project . --ttl 4h --install. Run entroly verify-claims and entroly simulate for local checks.

claude_desktop_config.json 
{"mcpServers":{"entroly":{"command":"entroly","args":[]}}}

Shown for Claude Desktop. Other clients may use a different file or key (VS Code uses "servers") — the configurator below converts it.

.vscode/mcp.json
{
  "servers": {
    "entroly": {
      "command": "entroly",
      "args": []
    }
  }
}

Goes in your project's .vscode/mcp.json (VS Code uses a "servers" key).

Terminal
claude mcp add entroly -- entroly

Run it in a terminal; replace any <…> placeholders with your own values first.

Generate and validate client config

Check that it works

Run `entroly verify-claims` to pass the packaged local self-test (no API key needed); in your MCP client, confirm the entroly server starts and exposes its context optimization, retrieval, receipt, and verification tools.

Troubleshooting

  1. Run entroly verify-claims to check installation, imports, compression, recovery, and engine mode
  2. Run entroly simulate or entroly perf to inspect local token estimates and optimizer latency
  3. Confirm the MCP client uses the stdio command entroly with no arguments
  4. For proxy mode, confirm the application base URL points to localhost:9377 and the upstream provider is configured
  5. If no provider-bound requests are observed, check whether the client bypasses the custom endpoint or is using MCP-only mode
  6. Verify that the receipt and recovery store still exist when exact recovery is required

Things to try

Once connected, you can ask your AI assistant things like:

  • Compress this API response to under 2000 tokens and give me a receipt of what was omitted.
  • Check whether this answer is supported by the supplied evidence and flag unsupported claims.
  • Run a local no-LLM token savings estimate on this repository.
  • Recover the exact omitted authentication-related code chunk that was compressed earlier.

Use cases

Select and compress repository or document context under a token budget
Compress repetitive terminal, test, and function output while retaining recovery handles
Create Context Receipts recording selected content, omissions, risks, and fingerprints
Recover omitted content exactly and replay committed context
Check answers against supplied evidence with WITNESS
Estimate local token reduction and measure provider-bound request value

Supported clients

Claude Desktop
Claude Code
Codex
OpenClaw
Cursor
Windsurf
VS Code MCP clients
Zed
GitHub Copilot CLIPartial support

Listed from the project's documentation, not tested by this site.

Overview

Entroly is a local Context OS and MCP server that discovers, ranks, deduplicates, and compresses tool output, logs, files, conversations, and RAG context before model requests. Budget-aware selection, content-addressed recovery, and Context Receipts preserve critical evidence while local verification, memory, caching, and cost-observation features provide additional control. The source claims up to 90% reduction in unnecessary tokens on measured workloads, but results depend on the workload and integration path.

Similar servers

Context7 80 · B

Upstash's official server providing up-to-date third-party library docs for AI coding assistants

★ 63.9k · Tools 2 Compare with this →
Compare with other MCP servers

Source revision a96745061bbe Data synced 2026-10-29 Read the FMRS scoring method