Entroly — Auditable Context Engineering for AI Agents

Know exactly what your AI agent saw.
Entroly creates replayable Context Commits: content-addressed proof of the evidence selected, omitted, and kept recoverable for each model request.

_{Integrates with Claude Code, Codex, OpenClaw, GitHub Copilot, Cursor, Aider, MCP clients, and custom provider applications. Choose the supported setup path for your client.}

_{Context selection + recoverable compression · receipts record what was used, omitted, and risky · local-first · Python with optional Rust acceleration · Node/WASM runtime}

_{Registry badges show distribution metadata. Evidence badges link to committed results with scope and caveats. Community and marketplace status are not treated as technical proof.}

pip install -U entroly && cd /your/repo && entroly verify-claims && entroly simulate

Get started · Proof videos · Proof · Integrations · What's inside · Architecture · For teams · Limitations

Deciding whether to star? Run the no-key proof first: entroly verify-claims && entroly simulate.
_{If it finds meaningful savings or gives you auditable receipts on your repo, star it so other agent builders can find it. If it does not, open an issue with the verification JSON.}

What it does

Entroly is an open-source, auditable context engineering control plane for AI agents. It decides what context to send, compresses selected evidence when useful, records what it left out, and produces a receipt you can inspect before trusting a hard multi-file answer.

OpenClaw runs the agents and conversations. Entroly controls, remembers, verifies, and proves the context those agents received. The same boundary applies to Claude Code and Codex: Entroly augments the agent you already use instead of becoming another chat client.

Most compression tools shrink whatever text the agent already chose. Entroly starts one step earlier: it chooses the highest-value evidence first, compresses only after selection, keeps originals recoverable, then verifies the answer against the evidence.

• Receipts - every selection run can explain selected chunks, omitted nearby evidence, dependency links, fingerprints, token ratio, and residual risks.
• Select - ranks your repo or document set, then sends the answer-relevant context under a token budget.
• Verify - WITNESS can check an answer against supplied evidence locally, without an additional model call. See the scoped benchmark under Proof.
• Route - sends easy, repeated tasks to a cheaper model and keeps the flagship for hard ones (opt-in, fail-closed).
• Cache-align - keeps the injected prefix byte-stable so provider prefix caches can keep hitting where terms and API shape allow it.
• Learn - adapts local ranking signals from recorded outcomes. No embeddings API or training job is required for that path.

Use it however you work: wrap your agent, run it as a proxy, plug it in as an MCP server, or import the library.

Why teams care

Product surface

Entroly ships as a full local runtime, not one proxy command:

Under the hood, the Python control plane has a pure-Python path and can use the optional Rust extension for supported operations. The separate Node runtime uses WASM. The implementation includes BM25, entropy scoring, SimHash dedup, dependency graphs, budgeted selection, caching, verification, and memory primitives; installed capabilities depend on the selected package and extras.

See the full code-derived map in docs/product-surface.md.

How it works (30 seconds)

your agent  ──►  Entroly (local)  ──►  LLM provider
                 │
                 ├─ rank the repo        (BM25 + entropy + dep-graph)
                 ├─ select under budget  (knapsack, reversible)
                 ├─ emit receipt         (included, omitted, risks)
                 ├─ cache-align prefix    (keep provider cache hot)
                 └─ verify the reply      (WITNESS hallucination guard)

Critical files go in full. Supporting files can become signatures. Other material can become a reference that can be expanded on demand. Exact recovery is available only while the corresponding receipt and recovery store are retained; deleting that state deletes Entroly's recovery path.

Proof in 30 seconds

Three short, reproducible checks show the value before asking you to trust the product. These are not mocked terminal recordings: each video is rendered from a checked-in command that verifies its source artifact before printing a number. Click an animation for the MP4, or run the command below it yourself.

1. Prove the installed path works—without an API key

The packaged verifier exercises import, compression, receipts, witness checks, recovery, proxy routing, and replay, then writes a machine-readable JSON report. It is an install smoke test—not a savings or model-quality benchmark.

entroly verify-claims

2. Prove tighter context can preserve more answers

Using Headroom today? Run Entroly against the same workload and compare answer retention, recoverability, state size, and context cost locally. The results below use the published Headroom 0.31.0 package as a versioned baseline; they are not a verdict on every Headroom or Entroly workload.

On the frozen 24-case holdout, Entroly answered 24/24 cases; the published Headroom 0.31.0 baseline answered 18/24, at 28.88% versus 42.97% effective context. This is a synthetic local Qwen2.5-1.5B test at temperature 0, not a universal product or model claim. The six discordant cases favored Entroly (exact McNemar p=0.03125).

python scripts/readme_proof.py model-recovery

Protocol and limitations · raw holdout artifact

3. Prove omitted evidence remains recoverable after restart

The prior v2 run tied at 66/66 and remains published. In the fresh-seed v3 Windows revalidation, Entroly recovered 66/66 payloads byte-exactly after restart; the published Headroom 0.31.0 baseline recovered 55/66. The raw artifact retains the incomplete worker and exact failure evidence. This is one reproducible run, not a universal durability claim.

python scripts/readme_proof.py restart-recovery

Protocol, prior tie, and limitations · current raw artifact · prior tie artifact

The animations, MP4s, static frames, source hashes, and commands are bound in the proof media manifest. Maintainers can rebuild them with python scripts/render_readme_proof_videos.py generate and reject stale media with python scripts/render_readme_proof_videos.py verify. Rebuilding requires Pillow, tiktoken, and FFmpeg; a missing frozen-tokenizer dependency fails with an actionable install command instead of weakening the artifact check.

Get started

The best first run is local and proof-driven. It should work before you connect an API key, proxy, paid model, or enterprise setup.

pip install -U entroly     # or: npm i -g entroly  ·  brew install juyterman1000/entroly/entroly

1. Prove the package works on your machine:

entroly verify-claims      # SDK import, indexing, optimization, exact recovery, engine mode
entroly simulate           # local no-LLM savings estimate on your current repo

2. Pick one integration path:

3. Best setup for Claude Code subscription users:

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

Claude Code stays your client. Entroly adds only the granted tools for compression, retrieval, receipts, and verification. Revoke access at any time with entroly attach revoke <grant-id> --uninstall; every tool call re-checks the grant, expiry, project, and scope.

4. One command — auto-detects your IDE, wraps your agent, opens the dashboard:

cd /your/repo && entroly go

5. Or wrap a specific agent:

entroly wrap claude     # Claude Code
entroly wrap cursor     # Cursor
entroly wrap codex      # Codex CLI
entroly wrap aider      # Aider

6. Or run the proxy — best for pay-as-you-go API keys and custom apps:

entroly proxy                                   # http://localhost:9377
ANTHROPIC_BASE_URL=http://localhost:9377     your-app
OPENAI_BASE_URL=http://localhost:9377/v1     your-app
GOOGLE_GEMINI_BASE_URL=http://localhost:9377/v1beta     your-app

7. Or measure it on your own repo first:

entroly demo            # before/after token + cost estimate
entroly simulate        # local no-LLM savings estimate
entroly perf            # local no-LLM savings + optimizer latency
entroly verify-claims   # runs the packaged self-test, writes a JSON report

Local-first: Entroly performs indexing and selection on-device. The selected prompt is sent only through the model integration you configure. Entroly does not send outbound analytics by default. Apache-2.0.

First-run success contract

Entroly should feel useful before you connect a paid model key:

• entroly verify-claims proves SDK import, local indexing, optimization, exact recovery, and native/pure-Python engine mode.
• entroly simulate shows the likely token reduction on your repo without making an LLM call.
• MCP setup works for Claude Code subscription users who do not want proxy/API-key mode.
• Proxy mode is available when you control the provider key and want transparent request optimization.
• npm/WASM is available for Node-first users, but Python remains the fullest CLI/SDK path.

If your repo is tiny or already under budget, Entroly should say so and pass through rather than invent fake savings.

Context Commits

A Context Commit is a portable JSON artifact for the exact context selected for an agent request. It binds the ordered selected text, omitted evidence, exact recovery data, engine/version identity, and optional parent lineage to one content-addressed ctx_... identifier.

entroly context-commit ./repo --query "Where is token rotation enforced?" \
  --budget 8000 --out context-commit.json
entroly context-commit --verify context-commit.json

from entroly import create_context_commit, replay_context, verify_context_commit

commit = create_context_commit(
    [("auth.py", open("auth.py", encoding="utf-8").read())],
    query="Where is token rotation enforced?",
    token_budget=8000,
)
assert verify_context_commit(commit).valid
exact_context = replay_context(commit)

The artifact is self-contained and therefore may contain source text in its recovery bundle. Keep it under the same access and retention policy as the source repository. Content addressing proves mutation, not signer identity; use Entroly's optional Ed25519 attestation and Merkle-log APIs when custody or operator identity matters. Contract and threat model.

Context Receipts

Receipt-producing selection workflows record what was used, what was omitted, why, and what risks remain. This is useful for hard multi-document work such as contracts, policies, addenda, code reviews, and audit evidence where a bare top-k result is not enough.

entroly ingest ./docs
entroly select --query "Does this contract have a change-of-control clause?" --budget 8000
entroly receipt .entroly/receipts/cr_example.json
entroly audit .entroly/session_chain.json
entroly explain --why-omitted chk_example --receipt .entroly/receipts/cr_example.json

The receipt JSON includes selected chunks, omitted relevant chunks, ranking reasons, dependency links, source fingerprints, token ratio, warnings, and a reproducibility hash. The Markdown report is designed for human review before a compressed context is trusted.

Implementation notes:

• Rust core (entroly-core/src/context_receipts.rs) handles deterministic ingestion, BM25-style ranking, dependency scans, selection, and hashes when the native wheel is available.
• Python control plane (entroly/context_receipts/) provides CLI wiring and a pure-Python fallback for source checkouts.
• The semantic/vector scorer and reranker are explicit extension points; the local MVP ships with lexical scoring and dependency heuristics, not a legal-accuracy guarantee.

Examples:

• Example receipt JSON
• Example Markdown report
• Limitations

Proof

Context Commit conformance (synthetic deterministic code fixtures, local, no model or network calls):

Reproduce: python -m benchmarks.context_commit_conformance. Raw JSON. These numbers measure artifact integrity, replay, and recovery on the committed fixtures; they do not measure model-answer quality or claim identical Python/Rust selection.

Context Efficiency Frontier research: Entroly is building a paired, model-neutral benchmark for the question token-savings tables cannot answer: does less context preserve real task quality? The preregistered protocol compares raw context, model-native compaction, Entroly, and their combination using provider-observed tokens, cost, latency, task success, evidence recall, and unsupported claims.

Read the preregistered protocol. No headline result will be claimed until the paired confidence bounds pass.

Matched token-cap active-context quality frontier (1.0.59 source candidate): across 60 frozen SQuAD v2 long-context RAG/tool-result trials, without invoking recovery from the published Headroom 0.31.0 baseline's CCR pointers, Entroly retained 95.0%, 93.3%, and 88.3% of accepted answers at the 2x, 4x, and 8x token caps. The published Headroom 0.31.0 baseline retained 1.7% at each cap. The paired two-sided exact McNemar tests pass at every point (p <= 4.45e-16). Entroly met all 180 caps; the baseline met the 2x/4x caps and exceeded the 8x cap, retaining 18.6% of tokens against the 12.5% maximum.

A separate eight-question, randomized local qwen2.5:1.5b guard at 4x scored raw context at 62.5% exact match, Entroly at 87.5%, and the published Headroom 0.31.0 baseline at 12.5%, with no errors. This small local-model sample is a veto guard, not the headline or a frontier-model claim.

Generated report · full auditable artifact · protocol and reproduction. Verify it with python -m benchmarks.compression_frontier verify benchmarks/results/compression_frontier.json.

_{Scope: extractive answer retention in structured SQuAD-v2 RAG results. The published Headroom 0.31.0 baseline's CCR pointers remain in its output, but retrieval recovery is not run; this measures immediately visible active context, not Headroom's end-to-end CCR workflow. Entroly is measured from the 1.0.59 source candidate; do not call this a released-package result until 1.0.59 is published. This does not establish superiority for every task, model, compressor, or production workload.}

Same-input compression gauntlet: on four deterministic agent-tool fixtures, current Entroly source (package version 1.0.59) and the published Headroom 0.31.0 baseline both retained 100% of the preregistered answer evidence. Under the shared o200k_base tokenizer, Entroly reduced weighted input tokens by 95.1% versus 31.4% for the baseline's public compress() pipeline with its documented agent-90 high-savings profile. Entroly compressed all four fixtures; the baseline compressed two and safely passed two through.

Generated report · raw inputs and outputs · protocol and reproduction. Verify the artifact with python -m benchmarks.compression_gauntlet verify benchmarks/results/compression_gauntlet.json.

_{This is a synthetic, no-model compression/evidence result. It does not measure downstream answer quality or establish neural/ML superiority. The Context Efficiency Frontier above is the required gate for a real-model claim.}

Cross-process recovery holdout: the preregistered six-writer test first exposed a serious Entroly failure (only 8/32 development payloads survived), which is preserved in the evidence. The original Entroly 1.0.59 holdout and the immutable v2 revalidation both recorded Entroly and the published Headroom 0.31.0 recovering 66/66 payloads. On the fresh-seed v3 revalidation of the current complete-line recovery implementation, Entroly again recovered 66/66 byte-exactly; the published baseline recovered 55/66. The raw artifact preserves the incomplete writer and exact failure evidence. Entroly was the participant that satisfied the frozen integrity gate on this run. That is a scoped concurrent-writer result; it does not establish universal recovery superiority.

On the current-implementation Windows/Python 3.10 revalidation, the published Headroom 0.31.0 baseline had lower successful store-call latency (1.786 ms versus 36.972 ms p50). Entroly had lower retrieval latency (0.165 ms versus 0.876 ms p50) and a smaller live state footprint (95,438 versus 1,354,888 bytes). The baseline used SQLite WAL with synchronous=NORMAL; Entroly fsynced its state file on each commit, so this is not a matched power-loss durability comparison. These are scoped workload measurements, not universal claims.

Frozen protocol and full result table | current v3 revalidation | prior v2 tie | original post-repair holdout | original failing artifact.

Quality-gated compression latency holdout: on the same four deterministic agent-tool fixtures and public entry points as the gauntlet, Entroly 1.0.59 source was 2.94x faster than the published Headroom 0.31.0 baseline for warm compressor calls (95% bootstrap CI 2.74x–3.13x) and 2.39x faster for product import plus the first call in a fresh process (1.89x–2.70x). Both systems completed every fixture, retained 100% of preregistered evidence, remained deterministic, and never inflated tokens.

Protocol, per-fixture timings, and limits | full holdout artifact | development artifact.

_{Scope: Windows/Python 3.10, synthetic local compression, 120 warm and 40 cold observations per participant. Cold excludes interpreter startup and includes product import plus first call. This is not provider latency, downstream answer quality, neural superiority, or universal product superiority.}

Model-triggered recovery holdout: after compression for one question, a different future audit question was revealed to a local qwen2.5:1.5b guard. On 24 frozen query-shift cases, raw context and Entroly both scored 24/24 exact; the published Headroom 0.31.0 baseline scored 18/24. All six paired differences favored Entroly (two-sided exact McNemar p = 0.03125). Entroly's mean effective context ratio, including recovery evidence on every triggered retry, was 28.88% versus 42.97% for the baseline.

Every Entroly row triggered retrieval and recovered a complete source-exact JSON object. The published baseline answered 18 rows from active context; its remaining six rows were scored under the frozen no-oracle retrieval rule. The raw artifact retains every response. The complete artifact passed the strengthened verifier with zero execution errors.

Protocol, rejected variants, reproduction, and limits | full holdout artifact | development artifact.

_{Scope: synthetic 48-record JSON audit logs, Windows/Python 3.10, 24 holdout cases, local Qwen2.5 1.5B Q4_K_M at temperature zero. The published Headroom 0.31.0 baseline uses its public compress() plus persistent CompressionStore contract; MCP transport is excluded. This is a scoped workflow result, not evidence about hosted frontier models, every agent workload, provider cost, or overall product superiority.}

PRISM-R neural research preview: a generic MiniLM encoder did not beat BM25 as a primary paragraph scorer (97.7% versus 99.0% held-out evidence recall), so Entroly rejects that neural-primary claim. A disagreement guard reached 99.3% while selecting 1.02 of 16 passages on average. In a separate 200-pair query-shift pilot at a nominal 25% active budget, PRISM-R retained 87.0% of current-query evidence versus lexical selection's 60.5%; when a different future question was revealed, exact receipt-backed rehydration raised its evidence retention from 9.0% to 90.5%. Active plus recovered context was 50.6% of the original.

Research design and prior art · held-out retrieval artifact · query-shift artifact.

_{These are offline exact-evidence pilots on frozen SQuAD v2 subsets, not downstream answer-quality, latency, production-savings, or general neural superiority claims. PRISM-R remains opt-in research code.}

The tables below link each reported number to its committed result. Treat them as evidence for those specific datasets, budgets, models, and commits—not as a guarantee for another repository. entroly simulate uses a local token estimate; use provider-observed usage for a billing or production claim.

Accuracy retention — does compression hurt answers? Measured with gpt-4o-mini; intervals are Wilson 95% CIs. Each row links its raw result file.

_{*pass-through: context already fit the budget, so Entroly left it unchanged. Reproduce: python benchmarks/run_readme_benchmarks.py (needs OPENAI_API_KEY). Full table + MMLU/TruthfulQA in DETAILS.}

Hallucination detection — committed HaluEval-QA balanced, both-answers-scored run:

Reproduce: python benchmarks/halueval_qa_faithful.py. Protocol and raw result. The shared-sample accuracies overlap within their reported uncertainty; this result does not establish superiority, general hallucination prevention, or production answer quality. The separate STAVE exploratory result is not used for the headline because it follows a different evaluation setup.

Works with your stack

OpenClaw: keep the evidence uniform compression drops

OpenClaw remains the conversation and agent runtime. Attach Entroly as its context control plane with entroly attach create --client openclaw --project . --ttl 4h --install.

The OpenClaw context-engine plugin keeps context assembly separate from provider routing. The same assembly path can be used with OpenAI, Anthropic, Gemini, Nemotron, OpenRouter, Ollama, and custom routes because OpenClaw owns routing and authentication. Its resolved prompt budget is authoritative; when an older or degraded host omits that value, Entroly can derive a conservative input ceiling from verified, operator-supplied, or explicitly discovered model metadata. Unknown and announced limits fail safely instead of being guessed. The receipt binds the discovery trust, registry digest, source, output reserve, and safety margin to the assembly decision. Entroly preserves opaque provider blocks and delegates /compact and overflow recovery back to OpenClaw.

The beta OpenClaw context engine scores older messages against the current request. Matching evidence is pinned verbatim when it fits a bounded reserve; lower-value history is compressed around it, and every decision is written to a local receipt.

In the committed synthetic, no-model control below, both strategies fit the same 1,800-token estimated budget. Uniform compression lost the exact old authentication instruction; evidence pinning retained it byte-for-byte.

Reproduce locally: python -m benchmarks.openclaw_evidence_pinning. Benchmark JSON · Plugin setup

_{Synthetic deterministic normalized multi-provider workload, 23,089 estimated source tokens, 11 messages, zero model calls. Token counts are estimates, not billed usage, and this result does not establish downstream model accuracy.}

entroly wrap <agent> picks the best integration for each tool — proxy env-wrap for CLIs, auto-merged mcp.json for MCP-aware IDEs, or a best-effort endpoint/config hint.

Wrap in one command: claude · cursor · codex · aider · gemini · windsurf · vscode · zed · cline · continue and 28 more.

As a library (LangChain, LlamaIndex, your own code):

from entroly import compress, compress_messages, optimize

compressed = compress(api_response, budget=2000)          # query-agnostic
messages   = compress_messages(messages, budget=30000)    # whole conversation
gentle     = compress_messages(messages, target_ratio=0.90)  # smooth relative operating point
context    = optimize(fragments, budget=8000, query="fix the login bug")  # task-conditioned

compress_messages infers the final user turn as the task and uses it to prioritize answer-relevant evidence in older context. target_ratio is based on Entroly's dependency-free token estimate; measure provider-observed tokens with the Context Efficiency Frontier before publishing a savings claim.

In CI — fail the build if a prompt blows the token budget:

- run: pip install entroly && entroly batch --budget 8000 --fail-over-budget

When to use it · when to skip

Great fit

• Large repos where the agent only sees a few files at a time
• Chatty, multi-turn agents (cache alignment compounds the savings)
• Anywhere you want answers checked against evidence before you trust them
• Teams trying to cut a real, growing AI bill

Skip it (it'll just pass through)

• Tiny repos or short prompts that already fit the budget
• Judgment-heavy tasks where you want the full flagship model every time

What's inside

Entroly exposes 19 local control mechanisms across input, inference, output, verification, and learning. The table describes their role and source location; it does not assign an unmeasured savings percentage to each mechanism.

pip install entroly            # core: MCP server + Python engine
pip install entroly[proxy]     # + HTTP proxy
pip install entroly[native]    # + Rust engine
pip install entroly[full]      # everything

npm install -g entroly         # WASM runtime, no Python needed
docker pull ghcr.io/juyterman1000/entroly:latest

cd entroly/entroly-core && cargo build --release --bin entroly-rs --features proxy
./target/release/entroly-rs proxy --upstream https://api.anthropic.com

WITNESS — check answers before you trust them

entroly witness --context-file evidence.txt --output-file answer.txt --mode strict
entroly proxy --witness strict --witness-profile rag    # suppress unsupported claims inline

Profiles tune false-positive behavior per workload (rag, qa, code fail closed; chat, summary warn). When WITNESS is enabled on a supported non-streaming proxy path, Entroly emits a certificate and the dashboard can show flagged claims, evidence snippets, and suppression counts. Optional offline DeBERTa NLI is enabled with ENTROLY_LOCAL_NLI=1; evaluate it on your workload before making an accuracy claim.

Why Entroly is different

The winning product is not the one that makes the prompt smallest. It is the one that helps the model do the best work for the fewest tokens.

Entroly is built around that trust contract: select the right evidence, compress supporting material, keep originals recoverable, emit a receipt, and verify the answer against the retained evidence.

The goal is same-quality or better model work at materially lower token cost.

Self-improving local runtime

Entroly has a guarded self-improvement loop. It is designed to learn from real outcomes without letting adaptation run wild.

This is the important distinction: Entroly does not just remember context. It can learn which context-selection strategies, routes, and skills actually produce successful work.

Compared to

Compressing a bad selection is still a bad selection. Entroly ranks first, then compresses — so the model gets structure, not just fewer tokens.

Docs & community

• Context control plane — model metadata, secure attachment, gateway recovery, and context-session UI guarantees.

• Architecture & full spec — Rust modules, 3-resolution compression, provenance, RAG comparison, SDK, LangChain.
• Product surface map — CLI, SDK, MCP, proxy, npm/WASM, verification, memory, and security surfaces.
• First-run trust guide — exactly what a new user should run before wiring a paid model.
• For teams — ROI, security, deployment one-pager.
• Limitations — where Entroly helps, where it passes through, and what it does not guarantee.
• Public evidence policy — claim tiers, benchmark scope, package links, and marketplace status.
• Cookbook — copy-paste recipes for common workflows.
• Discord Community · Discussions · Issues

Marketplace status

Marketplace pages are discovery surfaces, not release verification. The public LobeHub listing was still showing stale version and capability metadata in the latest recorded audit. Use the live page for current external status and the LobeHub score audit for the dated baseline. Until the listing matches the published package and passes external validation, install from PyPI or npm using the instructions above.

_{Apache-2.0 · local-first · no outbound analytics by default}

pip install entroly && entroly go