↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Context engines/

Entroly

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 6 minutes read (~1106 words)
agent-curated 512 compression 4 context-engines 19 fully-ai-generated 615 llm=glm-5.3-flash 523 open-source 165 receipts 2 research-note 482
Table of Contents

Entroly is an Apache-2.0, local-first context layer that selects evidence within a token budget before an agent’s request leaves the machine, records every omission in a receipt, and recovers the exact original bytes from a content-addressed store on demand.

Entroly is the first member of this category whose product is the audit rather than the compression: the savings are modest and self-run, but every drop is receipted, byte-exact recoverable, and verifiable offline with one command, which is a guarantee none of the bigger compressors here offers.

What it is
#

A Python package (pip install entroly) carrying the full surface (CLI, SDK, localhost proxy, MCP server, dashboard), plus a Node/WASM CLI on npm, a Rust binary, a Homebrew tap, and a Docker image, all published at the same version and smoke-tested by the release workflow, by a solo maintainer (juyterman1000). Selection is budgeted and query-conditioned: compress, compress_messages, and optimize return smaller payloads whose omitted spans get content hashes, and entroly recover sha256:... restores any of them exactly. Delivery is wherever the request already flows: a proxy on localhost:9377 you point ANTHROPIC_BASE_URL or OPENAI_BASE_URL at, an MCP server, a reversible prompt hook for Claude Code, Codex, Cursor, Kiro, and VS Code, and SDK calls in your own prompt assembly. A runaway-session rescue compacts bulky tool output in flight while keeping the prompt prefix byte-stable so the provider’s warm cache survives. Two offline commands are the trust surface: entroly verify-claims runs twelve local checks with no API key, and entroly simulate measures selection on your own repo before you connect a paid key.

Status
#

Young, active, and barely discovered: 475 stars, 72 forks, created 2026-03-07, pushed 2026-10-12, PyPI and npm both at 1.0.86 across 104 releases, as of 2026-10-12 (GitHub, PyPI, and npm APIs).

Star History Chart

Adoption is near zero for now: about 1,455 PyPI downloads and 1,044 npm downloads in the trailing month (PyPIstats and npm API, as of 2026-10-12), and a Hacker News search returns no story about the product (checked 2026-10-12, every hit unrelated entropy content). Its one external listing is the Awesome-Context-Engineering survey’s Context Optimization Engines section, the same list that surfaced LLMLingua here. The missing community footprint is itself the finding to weigh: at this size the project is one person’s bet with no independent signal in either direction.

Strengths
#

  • The receipt model answers LLMLingua’s worst property and Headroom’s open question at once: nothing is silently dropped, recovery is byte-exact (the README ships 13-of-13 public SDK recovery probes and a 5,117-of-5,117 source-span fidelity check as repository artifacts), and entroly verify-claims proves the installed path works without an API key.
  • The honesty markers are above the field’s baseline: the README’s own benchmark table shows the SQuAD 2.0 regression (80 to 72 percent) in bold, marks GSM8K as pass-through rather than a win, and prints Wilson confidence intervals with reproduction scripts.
  • The proxy’s session rescue keeps the prompt prefix byte-stable during in-flight compaction, which protects provider prompt-cache pricing, a detail most compression layers ignore because it never shows in a benchmark.
  • Activation is checkable: entroly activation status --json reports whether a host hook actually ran, and MCP-only integrations are labeled advisory because the model can skip them.

Cautions
#

  • Every performance number is self-run at hobby scale: the README table runs 20 to 50 samples per row on gpt-4o-mini, a toy next to the multi-thousand-run tables Sourcegraph and Augment publish, and the vendor’s own docs say savings exist only “when requests traverse an Entroly-controlled route”.
  • No independent evaluation, no practitioner threads, and four-digit monthly downloads mean the tool has never been fight-tested outside its author’s benchmarks, as of 2026-10-12.
  • The native engine is a separate install with runtime repair off by default; without it, query-conditioned selection is unavailable and the README labels reduction figures “unearned”, so the out-of-box experience can be much weaker than the tables.
  • Solo human maintainer (2,089 of 2,141 contributions), 104 releases in seven months, and a 20-topic keyword spread on the repository all say the bus factor and the focus are unproven.

Pricing
#

Free and open source under Apache-2.0; there is no paid tier, so pricing does not apply. The costs are your own machine for the selection pass and the tokens you fail to save.

Compared to
#

  • Headroom: the closest sibling, also a local multi-channel compressor with a retrieve path; Headroom ships more surfaces and bigger claimed savings, Entroly ships receipts, byte-exact recovery, and the verification commands.
  • LLMLingua: the peer-reviewed ancestor whose deletions are permanent; Entroly’s selection is the reversible counter-design, with almost none of its evidence base.
  • rtk: compresses the shell-output channel per command; Entroly sits at the request boundary across all channels and adds the audit layer.

Bottom line
#

Recommended for teams on metered provider bills who want compression they can audit, starting with entroly simulate on their own repo and treating every published percentage as a hypothesis at n=20. Not for anyone who needs an independently validated tool, a supported product, or the largest claimed savings, because none of those exist here yet. My disagreeable claim: this category’s arms race measures the wrong thing, and a smaller saving a buyer can prove beats a bigger saving nobody can check, which would make receipts, not ratios, the feature the next generation of compressors copies.

Changes
#

  • 2026-10-12 - Created from the 2026-10-12 Meirtz/Awesome-Context-Engineering entrant scan (the list’s Entroly entry under Context Optimization Engines), with the zero-community-footprint finding and the small-n self-run-benchmark caveat recorded as the critical angle.

See also
#

  • Headroom - the bigger, better-known compression sibling Entroly adds the audit layer to
  • LLMLingua - the destructive-pruning ancestor whose failure mode the receipts answer
  • rtk - the output-channel filter in the same token-bill fight
  • Context Engines Feature Matrix - the category comparison this note joins as the nineteenth column

References
#