openreview-cli

Privacy-first contract review automation CLI. Local-first, multi-agent AI that strips PII before any cloud call. Parse, review, and negotiate contracts through an AI Gateway spanning 17 providers all from the command line.

What is openreview-cli?

openreview-cli is a local-first, privacy-first contract review automation tool that runs entirely from the command line. It parses contract documents, strips personally identifiable information before any cloud call, reviews clauses through a multi-agent pipeline, and emits a structured memo. It runs fully locally by default, with cloud providers and privacy tiers available as explicit opt-in.

Why it matters

• PII stripped locally before cloud: the pipeline is fail-closed if page-level detection fails, the review halts rather than leak.
• Local-first by default: every slot defaults to local Ollama (qwen3 reasoning, nomic-embed-text embeddings). Sub in any local model per slot openreview config set gateway.reasoning.primary ollama/<model> or opt a slot into the cloud. Cloud providers are opt-in per slot.
• Multi-agent review pipeline: extraction → QA verification → citation grounding, not one monolithic prompt.
• 23 contract-type modes with bundled 3-position playbooks (Preferred / Acceptable / Walkaway).
• Dual human/agent interface: Typer CLI + Textual TUI for humans; Python API + JSON output for agents; plus an agent routing skill (skill/SKILL.md) for intent-to-command mapping.
• Spec-driven development: 33 specs, all tracked in specs/.

Agent skill

A project-local agent skill (skill/SKILL.md) maps an AI agent's intent to the right openreview command: review memos, bilateral comparison, retrieval over indexed chunks, negotiation analysis, gateway/playbook/PII management, privacy-tier config, and report export. It is a routing layer with 10 capabilities and 12 command-selection rules, and it disambiguates from the unrelated openreview.net academic platform.

Full routing rules and worked examples: skill/SKILL.md.

Quickstart

Install from PyPI (Python ≥ 3.12):

pip install openreview-cli

# One-time: configure the AI Gateway (local Ollama by default)
openreview gateway setup

# Review a contract
openreview precheck review contract.pdf

# Browse everything in the terminal UI (no args = launches the Textual TUI)
openreview

From source (contributors)

git clone https://github.com/mohamed-benoughidene/openreview-cli.git
cd openreview-cli && git submodule update --init && uv sync

# One-time: configure the AI Gateway (local Ollama by default)
uv run openreview gateway setup

# Review a contract
uv run openreview precheck review contract.pdf

# Browse everything in the terminal UI
uv run openreview

# Memo written to review_results/ (Markdown, JSON, or DOCX)

Measured results

Area	Value
Version	0.2.0 (alpha)
Tests	3,034 collected (10 markers)
Gateway	17 providers, 27 models
Contract modes	23
Startup	0.68 s median, ~43 MB RSS
Review accuracy	90.9% F1 (12 NDA clauses, claude-sonnet-4.6 via OpenRouter)
CUAD clause identification	100% sentence boundary recall (462 contracts, 4,034 queries)
PII detection	52.8% recall on 50 seeded contracts (spaCy en_core_web_lg)

Full methodology + raw numbers + measured-vs-unmeasured: BENCHMARKS.md.

What was measured

• Review accuracy: 90.9% F1 on 12 labeled NDA clauses (Claude Sonnet 4.6 via OpenRouter)
• Clause detection: 100% sentence boundary recall on CUAD v1 (462 real commercial contracts, 4,034 queries)
• PII detection: 52.8% recall on 50 seeded contracts (spaCy en_core_web_lg)
• Startup: 0.68 s median, ~43 MB RSS
• Honest gaps documented: what is NOT measured (e.g. retrieval on raw PDFs, end-to-end pipeline accuracy) is called out explicitly in BENCHMARKS.md

How it works

An async pipeline splits a contract into clauses, strips PII behind a fail-closed gate, and runs each clause through keyword match (no LLM), an extraction agent, a QA verification agent, and a citation-grounding discriminator before emitting a memo. The AI Gateway abstracts 17 providers through litellm with fallback, cost tracking, privacy tier routing, and API-key redaction. State lives in a single SQLite database, with per-document retrieval indexes alongside. → ARCHITECTURE.md

Architecture at a glance

• Pipeline: parse → PII strip (fail-closed) → clause split → per-clause multi-agent review (keyword match, extraction agent, QA verification, citation grounding) → structured memo
• AI Gateway: one litellm abstraction (chat / embed / rerank) across 17 providers, 27 bundled models, local Ollama by default
• Storage: single SQLite DB (reviews, cost logs, PII cache, playbooks, benchmarks) + per-document FTS5 retrieval indexes with RRF fusion
• Privacy: three tiers (Maximum / Balanced / Performance); nothing leaves the machine unless you opt in

Full architecture: ARCHITECTURE.md

Features by capability

Document processing

PyMuPDF page-by-page streaming parser (never loads the full PDF) plus python-docx for DOCX. Clause detection via 7 regex patterns + nupunkt sentence segmentation. Metadata extraction, corrupt/empty/password detection, non-English and tofu (broken glyph) detection. Streaming keeps memory bounded regardless of document length.

Privacy & PII

Presidio PII engine with spaCy en_core_web_lg plus 4 custom regex recognizers (AMOUNT, TAX_ID, ID_DOCUMENT, REG_NUMBER). Fail-closed by default: if page-level detection fails, the pipeline halts before any cloud call raises PartialProcessingError; opt out with --allow-partial-pii. Entities are replaced with [PARTY_A]-style placeholders; the reversible mapping is encrypted with Fernet (AES-128-CBC + HMAC, key via HKDF-SHA256) and stored chmod 600. Privacy tiers (Maximum / Balanced / Performance) control what leaves the machine.

AI Gateway

A single litellm abstraction: chat → completion, embed → embedding, rerank → rerank. 17 providers (openai, anthropic, google, openrouter, cohere, huggingface, deepseek, qwen, minimax, voyage, moonshot, mistral, zai, bedrock, azure, vertex, ollama). 27 bundled models; its default config points at local Ollama (qwen3:8b, qwen3:4b, nomic-embed-text) — bundled defaults, swappable for any local model. Per-slot fallback (2 retries, 60 s timeout), cost tracking to SQLite cost_logs (cents via litellm.completion_cost, non-fatal on error), configurable per-review/per-day limits (100¢/1,000¢ defaults, warn-only). API-key pattern redaction on all log output. Streaming with 15 s connect / 45 s idle timeouts.

Multi-agent review

Per-clause pipeline: keyword category match (no LLM) → extraction agent (LLM, outputs position + confidence + citation) → QA verification agent (LLM, agree/disagree/uncertain verdict, amber flag) → citation grounding discriminator (LLM, claim-vs-source verification, strict/lenient modes). 24 bundled playbooks across 23 modes. 3-position model: Preferred / Acceptable / Walkaway. 3-color confidence output: Green / Amber / Red with configurable threshold.

Analysis tools

• Game-theoretic negotiation (openreview negotiate): pure local NumPy Nash (support enumeration), QRE (logit fixed-point), Level-k (k ≤ 3) solvers. No LLM calls.
• Bilateral comparison (precheck compare): experimental RCBSF 5-dimension divergence detection, 3-tier heading alignment, ≤64% F1 ceiling (documented).
• Contract graph (openreview graph): directed clause graph, 0–100 health score from 5 structural metrics, optional legal-bert + HDBSCAN clustering, ASCII tree view, graph diff.

Storage & retrieval

Single SQLite database (19 tables: reviews, cost_logs, PII cache, playbooks, benchmarks, graph data, recovery state). Per-document retrieval indexes in separate SQLite files with FTS5 (BM25 unicode61, prefix 2–3) + dense embeddings (brute-force cosine scan, no vector DB honest limitation) + RRF fusion (k=60). Reranker present but disabled by default (degrades legal text); opt-in with --rerank.

Terminal UI

A full Textual app (src/openreview_cli/tui/) for humans who prefer browsing over flags: review documents, inspect the AI Gateway, and configure the tool — without typing subcommands. Launch with a bare openreview. The CLI and the TUI share the same pipeline, so a review started in one is identical in the other.

Contract-type modes

Mode	What it reviews
precheck	Non-Disclosure Agreement (NDA)
licensecheck	SaaS/software license agreement
leasecheck	Commercial lease agreement
privacycheck	Data Processing Agreement (DPA)
privacycheck_v2	Data Processing Agreement (v2)
dealcheck	Vendor/service agreement
hirecheck	Employment agreement
indemnitycheck	Indemnification agreement
consultcheck	Consulting services agreement
workcheck	Independent contractor/work-for-hire agreement
loicheck	Letter of intent or MOU
subcheck	Subcontractor agreement
settlementcheck	Settlement/release agreement
settlementcheck_v2	Complex settlement/release agreement (v2)
assetcheck	Asset transfer/assignment agreement
buycheck	Asset purchase/business acquisition agreement
engagecheck	Professional services engagement letter
guaranteecheck	Personal guarantee/suretyship agreement
loancheck	Loan agreement/promissory note
franchisecheck	Franchise agreement or franchise disclosure document
opcheck	Operating Agreement (LLC governance document)
partnercheck	General or limited partnership agreement
sponsorcheck	Sponsorship agreement
distrocheck	Distribution or reseller agreement

Each mode has a bundled 3-position playbook and mode-specific prompt vocabulary.

Privacy tiers

Tier	PII processing	Reasoning	Embedding	Description
Maximum	All local (fail-closed)	Local (Ollama)	Local	Nothing leaves the machine
Balanced (default)	All local (fail-closed)	Cloud (OpenRouter, OpenAI…)	Local	PII stripped, reasoning in cloud
Performance	All local (fail-closed)	Cloud	Cloud	PII stripped, all inference in cloud

Configure with openreview config set privacy.tier maximum|balanced|performance or env var OPENREVIEW_PRIVACY_TIER.

Essential commands

openreview      → launch the Textual terminal UI (browse, review, configure)
parse          → inspect clauses in a PDF/DOCX
precheck review → run full review (parse → PII strip → extract → QA → memo)
gateway setup  → configure AI providers (one-time)
gateway test   → verify each slot can connect
negotiate      → run game-theoretic negotiation analysis
export         → batch-export saved review reports

Run openreview --help for all 77 subcommands. No args launches the Textual TUI.

FAQ

Does openreview-cli send contract text to the cloud?

No not by default. PII is stripped locally by Presidio before any cloud call. The pipeline is fail-closed: if a page-level detection fails, the review halts immediately (use --allow-partial-pii to opt out). Cloud providers are opt-in per slot.

Can it run fully offline?

Yes. The default gateway configuration points at local Ollama models (qwen3:8b reasoning, qwen3:4b extraction, nomic-embed-text embeddings). Set the privacy tier to Maximum to guarantee nothing leaves the machine.

How accurate is it?

Measured against public benchmarks: 90.9% F1 on contract clause review (12 labeled NDA clauses through Claude Sonnet 4.6), 100% sentence boundary recall on CUAD v1 (462 real commercial contracts), 52.8% PII recall on 50 seeded contracts. All numbers, methodology, and unmeasured gaps in BENCHMARKS.md.

What can AI agents do with it?

Every module is independently importable as a Python library (from openreview_cli.parsing.stream import parse_document, from openreview_cli.review.extraction import extract_clause). The CLI supports --format json and --output for structured output. The benchmark runner (BenchmarkRunner) and pipeline (run_review) are Python-callable with typed return values. For intent-to-command routing (not raw library calls), use the agent skill at skill/SKILL.md.

What file formats are supported?

PDF (PyMuPDF, page-by-page streaming, password-protected, corrupt/empty detection) and DOCX (python-docx, track-changes, images, flat documents).

Is openreview-cli free?

Open source under AGPL-3.0 (LICENSE). A commercial license option is available (COMMERCIAL_LICENSE.md).

Tech stack / Licenses / Status

Python 3.12 · Typer CLI · Textual TUI · Presidio (PII) · litellm (gateway) · PyMuPDF / python-docx (parsing) · SQLite + FTS5 (storage, BM25) · numpy / scikit-learn (embeddings, solvers) · torch / transformers (CPU-only) · nupunkt · pydantic · cryptography · rich · jinja2 · httpx · pyyaml. Dev: pytest, mypy (strict), ruff.

AGPL-3.0-only, with a commercial license option (see LICENSE and COMMERCIAL_LICENSE.md).

Alpha. Measured performance, accuracy, and methodology in BENCHMARKS.md.

Architecture · Benchmarks · Issues · Discussions