Entroly

Entroly — Context Assurance That Helps Lower AI Costs

Use less unnecessary AI context without rebuilding how you work.
Entroly selects useful evidence, removes avoidable repetition, keeps important originals recoverable, and records what your AI received.

Entroly is a local-first Context Assurance layer for developers, teams, and AI applications. It works through supported proxy, wrapper, plugin, SDK, and MCP paths with Claude Code, Codex, OpenClaw, Hermes Agent, OpenCode, GitHub Copilot, Cursor, Aider, local models, and OpenAI/Anthropic-compatible applications.

A small one-time setup is required; no agent-architecture rewrite.

What Entroly does

AI applications often send repeated files, long histories, logs, and low-value material to a model. Entroly can:

select useful evidence under an explicit budget;
remove duplicate or low-value context;
preserve omitted originals through content-addressed recovery handles;
create Context Receipts showing what was included, omitted, and risky;
verify whether output claims are supported by supplied evidence;
report provider-observed usage separately from local-only reductions;
work across cloud and local models without locking you to one agent runtime.

Savings and quality depend on the workload, provider, model, budget, and integration. Entroly does not promise a universal compression percentage or guaranteed bill reduction; actual savings depend on observable request boundaries.

Install

pip install -U entroly

Check the local path before connecting a paid model:

entroly verify-claims
entroly simulate
entroly value

verify-claims checks installation, receipts, recovery, and verification.
simulate estimates context reduction locally without making a model call.
value separates observed provider-bound usage from local-only reductions.

Connect a supported AI tool

Start the local proxy and register MCP for a detected project client:

cd /your/repo
entroly go

This does not redirect model traffic automatically. Point an API client at the displayed proxy URL, or use a supported host hook. After a real request, check entroly value --json for provider-observed traffic or entroly activation status --json for a recent host-hook receipt. MCP registration alone does not prove the agent used Entroly.

Examples:

# Claude Code
entroly attach create --client claude --project . --ttl 4h --install

# Codex
entroly attach create --client codex --project . --ttl 4h --install

# OpenClaw
entroly attach create --client openclaw --project . --ttl 4h --install

# Pay-as-you-go API or custom application
entroly proxy

For an MCP client, register the installed entroly command with no arguments:

entroly

Package-runner alternatives:

uvx --from entroly entroly
npx -y entroly-mcp

For a one-shot complete CLI without a global install, either Python tool runner works:

uvx --from entroly entroly --help
pipx run --spec entroly entroly --help

Claude Code MCP registration:

claude mcp add entroly -- uvx --from entroly entroly

Generic MCP configuration:

{
  "mcpServers": {
    "entroly": {
      "command": "uvx",
      "args": ["--from", "entroly", "entroly"]
    }
  }
}

How it works

Files, documents, history, and tool output
                  ↓
       Entroly Context Assurance
       • rank useful evidence
       • select under budget
       • preserve exact originals
       • create a receipt
       • optionally verify output
                  ↓
        Your existing AI model

Local indexing, selection, receipts, and recovery storage do not require an Entroly cloud service. The selected prompt still goes to the provider configured by your application. No outbound analytics are enabled by default.

Honest cost reporting

Entroly keeps different evidence classes separate:

Path	What can be reported
Provider-bound proxy request	Observed pre/post input tokens and modeled input-cost avoidance with pricing provenance
SDK, MCP, plugin, or npm operation	Local context reduction; `$0` claimed when provider delivery is not observable
Fixed-price subscription	Context efficiency where supported; the subscription price may not change
Unknown or legacy history	Preserved operational history; excluded from savings claims

Modeled provider cost avoidance is not a provider invoice.

Technical capabilities

budgeted evidence selection;
recoverable compression;
Context Receipts and Context Commits;
strict ccr:<24-hex> exact-recovery handles;
hash-only full-content retrieval with no semantic-query fallback;
WITNESS evidence-support verification;
pure-Python runtime with optional Rust acceleration;
separate Node/WASM runtime;
MCP, proxy, CLI, SDK, CI, OpenClaw, Hermes Agent, and OpenCode integrations;
local-first operation and fail-open adapter behavior.

WITNESS evaluates support against supplied evidence; it does not establish universal truth.

Everyday users

A future desktop experience, Entroly Simple Mode, is specified for people who do not want to learn about tokens, proxies, MCP, JSON, or agent architecture. It is not shipped yet, and Entroly does not claim a one-click consumer experience today.

Links

AI cost optimization: https://juyterman1000.github.io/entroly/ai-cost-optimization.html
Repository: https://github.com/juyterman1000/entroly
Documentation: https://juyterman1000.github.io/entroly/
Agent integrations: https://juyterman1000.github.io/entroly/agent-integrations.html
PyPI: https://pypi.org/project/entroly/
npm runtime: https://www.npmjs.com/package/entroly
npm MCP bridge: https://www.npmjs.com/package/entroly-mcp
Evidence policy: https://github.com/juyterman1000/entroly/blob/main/docs/public-evidence.md
Limitations: https://github.com/juyterman1000/entroly/blob/main/docs/limitations.md

mcp-name: io.github.juyterman1000/entroly

Apache-2.0 licensed.

agent-memory ai-agents ai-coding-agents ai-cost-optimization claude-code codex content-addressed-recovery context-assurance context-compression context-control-plane context-engineering context-management context-optimization context-receipts context-window cursor github-copilot hallucination-detection hermes-agent llm llm-cost-optimization llm-proxy local-models mcp mcp-server model-context-protocol openclaw opencode prompt-compression reduce-ai-costs token-optimization token-reduction

Metadata

Release files for entroly 1.0.87

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for entroly 1.0.87
File	Size	Uploaded
entroly-1.0.87.tar.gz	9.7 MB	Oct 13, 2026	Details

Built distribution (wheel)

Table of built distributions (wheels) for entroly 1.0.87
File	Interpreter	ABI	Platform	Reset
entroly-1.0.87-py3-none-any.whl 2.2 MB Oct 13, 2026	Python 3	none	any	Details

Total release size: 11.8 MB

Release files / entroly-1.0.87.tar.gz

Download URL	entroly-1.0.87.tar.gz
Size	9.7 MB
Tags	Source
SHA-256 checksum How to use checksums	`b03123990060764f389c960a94439a6401c4be6955b1476f62546f4d41c02017`
BLAKE2b-256 checksum How to use checksums	`724626ea5bdb591b9b2f5a994dd77ad136e1e14f8cd887802a3a550d6c0117e6`
Upload date	Oct 13, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.13.14`

Release files / entroly-1.0.87-py3-none-any.whl

Download URL	entroly-1.0.87-py3-none-any.whl
Size	2.2 MB
Tags	Python 3
SHA-256 checksum How to use checksums	`6b7a49a0676c109cd0717b67ba4d8cc1c385f8624cba7476eadc3789d5d49f19`
BLAKE2b-256 checksum How to use checksums	`ed492f1bc61a61aec59f4b2b32102f8d79631b61fcd8d5a40fb0c4657cfb98a3`
Upload date	Oct 13, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.13.14`