entroly 1.0.65

Open-source Context OS for AI agents: auditable context engineering, compression, recovery, receipts, and verification.

Entroly — The Open-Source Context OS for AI Agents

Keep your agent. Give it a Context OS.
The observability, governance, and decision layer for AI context.

Entroly is an open-source Context OS for AI agents: auditable context engineering, recoverable compression, memory, verification, provider controls, receipts, security, and guarded outcome learning in one local layer.

Install

pip install -U entroly

Run the local, no-key verification path:

entroly verify-claims
entroly simulate
entroly value

entroly value keeps provider-bound cost avoidance separate from SDK, MCP, and npm reductions. Local-only operations report tokens reduced with $0 claimed; modeled provider cost avoidance includes pricing provenance and is not a provider invoice.

MCP server

For an MCP client, register the installed entroly command with no arguments. When an MCP client launches it with a stdio pipe, Entroly starts the installed Python server directly:

entroly

Or register a package runner, also with no serve argument:

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

entroly serve is a different deployment path: it uses the Entroly Docker image by default. For the installed Python runtime in an interactive shell, use ENTROLY_NO_DOCKER=1 entroly serve on macOS/Linux or set ENTROLY_NO_DOCKER=1 in the client environment.

Entroly works with Claude Code, Codex, OpenClaw, GitHub Copilot in VS Code, Cursor, Windsurf, Cline, Continue, Zed, and other MCP-compatible clients.

GitHub Copilot / VS Code

Create .vscode/mcp.json:

{
  "servers": {
    "entroly": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--from", "entroly", "entroly"]
    }
  }
}

External MCP galleries can lag a release. Direct stdio registration above is the canonical setup; confirm a gallery entry's package version and validation status before relying on it.

Claude Code

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

Generic MCP configuration

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

What Entroly adds

• Context selection under explicit token and cost budgets
• Context Commits linking selected, omitted, and recoverable evidence
• Context Receipts for replay, audit, and omission explanations
• Exact recovery of compressed fragments through stable handles
• Proof-guided recovery that verifies drafts, recovers exact omitted evidence, and stops under declared round/token bounds; local prepare and advance operations never call a provider
• Local verification through WITNESS and receipt checks
• Context Check coverage evidence for changed files and CI risk gates
• Verified model-based dreaming (experimental, opt-in): real transitions train the model, synthetic rollouts only rank experiments, and real holdout evidence remains mandatory for promotion
• Pure-Python base runtime, optional Rust acceleration, and a separate npm/WASM runtime
• Local-first operation with no outbound analytics by default

Prepare a restart-safe model request without a provider call:

entroly proof prepare ./docs --query "What evidence supports this answer?" \
  --budget 8000 --idempotency-key request-001

The caller sends the returned request through its existing model route and returns the draft with entroly proof advance. See the repository's proof-guided protocol guide for MCP, proxy, and opt-in OpenClaw automation.

Links

• Repository: https://github.com/juyterman1000/entroly
• Documentation: https://juyterman1000.github.io/entroly/docs/index.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
• Public evidence policy: https://github.com/juyterman1000/entroly/blob/main/docs/public-evidence.md

MCP Registry identity

mcp-name: io.github.juyterman1000/entroly

Apache-2.0 licensed.