Patron

License: AGPL v3 Tests AI Act RODO Stack MCP Node

A local-first, self-hosted AI agent for a law firm. A zero-cloud, single-user desktop application (Electron): local SQLite by default (ADR-0053), 20 bundled connectors to Polish, EU and national law (SAOS / NSA / ISAP / KRS / EUREKA / EUR-Lex / EU-Compliance), a hash-chained audit trail (AI Act art. 12), bring-your-own-model (Gemini / Claude / local Ollama / OpenRouter). A server mode (Postgres + MinIO) remains available as an alternative.

Patron is a fork of Mike (a document-centric legal assistant, AGPL-3.0). The Patron shell inherits AGPL-3.0 as a derivative work. It adds Polish localization, a Polish legal stack, and the compliance requirements a law firm needs. The full rules live in governance/CONSTITUTION.md.

Everything below ran end-to-end on 2026年08月23日 against a real model on synthetic contract fixtures (npm run smoke:surfaces, smoke:desktop, packaged app e2e:smoke); nothing here is a roadmap item.

• Chat over the case files with a colour badge next to every citation from your own documents: green (verbatim), yellow (paraphrase), red (the source does not support the claim). A local paraphrase judge, never a cloud call, decides the colour.
• Case law, legislation and tax practice while answering - 20 MCP connectors ship in the installer: 7 for Polish and EU sources (SAOS, NSA, ISAP, KRS, EUREKA, EUR-Lex/CJEU) queried live, plus 13 national ELI connectors (DE, FR, IT, ES, NL, SE, AT, FI, IE, LU, BR, GB, US) that follow the European Legislation Identifier. On top of that an offline EU compliance corpus (GDPR, AI Act, DORA, NIS2, eIDAS 2.0, CRA) that works without internet. Connector answers carry source, signature or identifier, date and URL (measured on SAOS and EUREKA; the same MCS contract applies to the rest).
• A table from a batch of contracts (Tabular Review): you define the columns as questions, Patron fills one cell per document with the answer, a quote from the source page and a confidence flag (green / yellow / red / grey).
• Editing documents as a lawyer does - you ask for a change and review it as Word tracked changes, then accept or reject. A file edited in Word comes back into Patron with its changes intact.
• Drafting - generate a new .docx from the chat, or run a whole pleading through a review, devil's advocate and language pass before it leaves the firm.
• Workflows - saved prompts and column sets you reuse across cases.
• Human-in-the-loop for agent writes (optional, ADR-0137): edits and generated files can be staged as approval cards; nothing is written until a human approves.
• An audit trail you can hand to a regulator - every model interaction hash-chained (AI Act art. 12) with a Merkle root, exportable as a ZIP that ships its own verifier (a browser page and a Python script), so a court, the DPA or the client can check it without this repository (ADR-0142).
• PII does not leave the machine by accident - names, companies, PESEL/NIP/REGON are masked before any cloud model; a per-case cloud consent and an egress guard sit in front of every outbound call.
• Bring your own model - Gemini, Claude, OpenRouter, or a local Ollama model, chosen in settings or per conversation.
• 9 installer editions from one code line: PL, EN (EU-first), US, GB, BR, IT, DE, ES, FR - each with its home jurisdiction connector on by default. Downloads: matematicsolutions.com/pobierz.

Two limits we say out loud. The memory and the knowledge graph work underneath: there is no screen yet where you can review or correct what Patron has remembered - you ask for it in the chat. The Windows installer is not code-signed, so SmartScreen warns on first run.

Written for the lawyer who has just downloaded the installer, not for the person reading this file. Kept in one place and linked from here rather than copied, so the two cannot drift apart:

• What Patron can do - every capability with the problem it answers and a link to the architecture decision behind it.
• Step-by-step manual - installation, connecting a model, your first matter, and the pitfalls that actually occur.
• Audit trail documentation - how the hash chain and the Merkle tree work, and what the trail does not prove.
• docs/SAMOUCZEK_EN.md - the same manual as markdown. Each installer edition ships the tutorial for its own language; this repository holds all nine.

• frontend/ - Next.js application
• backend/ - Express API, MCP client, audit trail, tool dispatch
• backend/src/lib/input-security/ - a local, deterministic scan of incoming documents (prompt injection / hidden PDF actions / obfuscation) before they reach the model or RAG (ADR-0019/0020)
• backend/src/lib/mcp-security/ - a local, deterministic scan of MCP connector definitions (typosquatting / description drift / hidden instructions / tool poisoning) BEFORE they are loaded into the MCP contract (ADR-0025/0028)
• backend/schema.sql - Postgres schema (server mode, Supabase-compatible); desktop mode uses local SQLite (ADR-0053)
• governance/ - Patron AI Constitution + Implementation Playbook + ADRs
• deploy/ - deployment runbook (docker-compose)
• scripts/bundle-mcp.cjs - bundler that packs the 7 MCP servers into the backend image

The seven connectors above ship inside the desktop installer. They are part of a larger fleet: 44 open MCP connectors under github.com/matematicsolutions, one repo per source, each reading an official government API or gazette. Language editions of Patron pair the same AGPL shell with the connectors for that jurisdiction. The full catalog with install commands lives at MateMatic Boutique.

Poland (MIT / Apache-2.0):

European Union (TypeScript, MIT):

France, offline (JavaScript, MIT):

National legislation, 33 jurisdictions (Python, Apache-2.0) - the xx-eli-mcp family, each reading the official national source with ELI-style citable references:

Austria · Australia · Belgium · Brazil · Canada · Chile · Colombia · Croatia · Czechia · Denmark · Finland · France · Germany · Hungary · Ireland · Israel · Italy · Japan · Lithuania · Luxembourg · Malaysia · Malta · Netherlands · Pakistan · Romania · Singapore · Slovakia · Spain · Sweden · Switzerland · Turkey · United Kingdom · United States

Cross-jurisdiction (Python, Apache-2.0):

Full runbook: deploy/README.md . The short version:

# 1. Clone the 8 repos (patron + 7 mcp-*)
git clone matematicsolutions/patron && cd patron
for d in mcp-saos mcp-nsa mcp-isap mcp-krs mcp-eureka mcp-eu-sparql mcp-eu-compliance; do
 (cd .. && git clone matematicsolutions/$d && cd $d && npm install && npm run build)
done
# 2. Bundle MCP into the backend image
node scripts/bundle-mcp.cjs
# 3. Configure secrets
cp .env.docker.example .env.docker
nano .env.docker
# 4. Up
docker compose --env-file .env.docker up -d

This requires a separately provisioned Supabase + MinIO (a separate stack). See the runbook.

• Patron AI Constitution v1.7.2 - 9 principles, product boundaries, roles (Administrator / Operator / Inspector), audit, and evolution. Mapped to AI Act art. 12, RODO art. 5/25/30/32, and professional ethics. Art. 5 covers input-document control.
• Implementation Playbook - a 6-8 week step-by-step rollout with a RACI matrix.
• ADRs - Architecture Decision Records (0001-0145), including 0001 hash-chain, 0002 dual-license, 0019 input-document scan, 0020 wiring into ingest.

The firm reads and signs the Constitution v1.7.2 before deployment (the signature section is at the end of the file).

Patron is the reference implementation of the open citation standard MateMatic Connector Standard (MCS) v0.1: the structuredContent.citations contract (source_id / url / exact_quote / locator / confidence) plus a 3-color credibility gradient (verbatim / paraphrase / unverified = existence / content / fragment) plus a conformance test (citation roundtrip). Any legal-source connector that satisfies MCS plugs into the citation-verification layer with no changes.

The stack is dual-licensed (see ADR-0002):

• patron (this repo, the shell) - AGPL-3.0-only (LICENSE + NOTICE)
• mcp-saos, mcp-nsa, mcp-isap, mcp-krs, mcp-eu-sparql, mcp-eu-compliance - MIT

A self-hosting firm can use, modify, and distribute Patron inside its organization with no extra obligations. A competitor that offers Patron as SaaS to third parties must open its modifications.

Patron is a fork of Mike (AGPL-3.0, (c) 2025 Will Chen). Full attribution: NOTICE.

The rest of this README covers the local (development) setup. For a production deployment, use deploy/README.md (Docker).

• frontend/ - Next.js application
• backend/ - Express API, Supabase access, document processing, and database schema
• backend/schema.sql - Supabase schema for fresh databases
• backend/migrations/ - incremental database updates for existing deployments

• Node.js 20 or newer
• npm
• git
• A Supabase project
• A Cloudflare R2 bucket, MinIO bucket, or another S3-compatible bucket
• At least one supported model provider API key: Anthropic, Google Gemini, or OpenAI
• LibreOffice installed locally if you need DOC/DOCX to PDF conversion

For a new Supabase database, open the Supabase SQL editor and run:

-- copy and run the contents of:
-- backend/schema.sql

The schema file is based on supabase-migration.sql and folds in the later files in backend/migrations/.

For an existing database, do not run the full schema file over production data. Apply the incremental files in backend/migrations/ instead.

Create local env files:

touch backend/.env
touch frontend/.env.local

Create backend/.env:

PORT=3001
FRONTEND_URL=http://localhost:3000
DOWNLOAD_SIGNING_SECRET=replace-with-a-random-32-byte-hex-string
SUPABASE_URL=https://your-project.supabase.co
SUPABASE_SECRET_KEY=your-supabase-service-role-key
R2_ENDPOINT_URL=https://your-account-id.r2.cloudflarestorage.com
R2_ACCESS_KEY_ID=your-r2-access-key
R2_SECRET_ACCESS_KEY=your-r2-secret-key
R2_BUCKET_NAME=patron
GEMINI_API_KEY=your-gemini-key
ANTHROPIC_API_KEY=your-anthropic-key
OPENAI_API_KEY=your-openai-key
RESEND_API_KEY=your-resend-key
USER_API_KEYS_ENCRYPTION_SECRET=your-long-random-secret

Create frontend/.env.local:

NEXT_PUBLIC_SUPABASE_URL=https://your-project.supabase.co
NEXT_PUBLIC_SUPABASE_PUBLISHABLE_DEFAULT_KEY=your-supabase-anon-key
NEXT_PUBLIC_API_BASE_URL=http://localhost:3001

Supabase values come from the project dashboard. Use the project URL for SUPABASE_URL / NEXT_PUBLIC_SUPABASE_URL, the service role key for the backend SUPABASE_SECRET_KEY, and the anon/public key for NEXT_PUBLIC_SUPABASE_PUBLISHABLE_DEFAULT_KEY. If your Supabase project shows multiple key formats, use the legacy JWT-style anon and service role keys expected by the Supabase client libraries.

Provider keys are only needed for the models and email features you plan to use. Model provider keys can be configured in backend/.env for the whole instance, or per user in Account > Models & API Keys. If a provider key is present in backend/.env, that provider is available by default and the matching browser API key field is read-only.

Install each app package:

npm install --prefix backend
npm install --prefix frontend

Start the backend:

npm run dev --prefix backend

Start the main app:

npm run dev --prefix frontend

Open http://localhost:3000.

1. Sign up in the app.
2. If you did not set provider keys in backend/.env, open Account > Models & API Keys and add an Anthropic, Gemini, or OpenAI API key.
3. Create or open a project and start chatting with documents.

Sign-up confirmation email never arrives. Confirmation emails are sent by Supabase Auth, not by Mike. For local development, the simplest fix is to disable email confirmation in Supabase > Authentication > Providers > Email. For production, configure custom SMTP in Supabase; the built-in mailer is heavily rate-limited and may be restricted on newer projects.

The model picker shows a missing-key warning. Add a key for that provider in Account > Models & API Keys, or configure the provider key in backend/.env and restart the backend.

DOC or DOCX conversion fails. Install LibreOffice locally and restart the backend so document conversion commands are available on the process path.

npm run build --prefix backend
npm run build --prefix frontend
npm run lint --prefix frontend