🏺 Arke

What this is

Arke is a self-hosted Python RAG for stress-testing UK competition law skeletons against CAT, EWHC, COA, UKSC, CJEU, GC authorities. Install it as an MCP server for your AI assistant, or run it on-prem connected to your firm's archive. Try a public demo at arke.legal.

Risk assessment

Privilege-safe by design

The on-prem mode has no external calls, no third-party APIs, no cloud storage, no sidecars, no database. Arke is a monolithic Python server with local inference you control privately.

Source data integrity

Arke never edits the source documents, only mirrors them, read-only. When documents are added, removed, or modified by the user, Arke will pull the change automatically.

Output data integrity

The system output is a verbatim mosaic of the source data - a bouquet of common-law authority citations. Thus, hallucinations are impossible.

---

AI-curated documentation

Everything below is curated by AI for AI. Paste this README into Claude / ChatGPT / Cursor and ask for install or configuration help.

Install (~2 min, MCP server)

Arke ships a stdio MCP server (arke-mcp) that plugs into Claude Desktop, Claude Code, or Cursor. It exposes two tools to your AI assistant:

• stress(argument) — full pipeline. Send a draft skeleton paragraph; receive verbatim cited authorities and an adversarial mosaic. 30-60s.
• search(query) — retrieval only. Top-K authorities with neutral citation, court, date, and snippet. Sub-second.

Both tools query a local UK CAT + EU competition law corpus (~3,400 documents, ~5GB on disk, auto-downloaded on first server boot).

Prerequisites

• Python 3.12+
• uv — Python package manager
• Inference backend — either an OpenAI API key (out-of-the-box default; same trust boundary as your existing Claude Desktop / ChatGPT usage) or a local inference server (llama-cpp-server or similar speaking the /v1/embeddings and /v1/chat/completions shape). See Configuration below to switch.
• One of: Claude Desktop, Claude Code, Cursor

1. Clone and install

git clone https://github.com/padolgot/arke.git
cd arke
./scripts/install.sh

The installer prompts you for your OpenAI key, writes a local .env, and runs uv sync. Your key never leaves your machine.

2. Start the local server

uv run arke-server &

First boot downloads the corpus (~5GB, 1-5 minutes). Subsequent boots are instant. Leave arke-server running in the background while you use your AI assistant.

3. Wire up to your AI assistant

Claude Desktop

Settings → Developer → Edit Config. Add to mcpServers:

{
  "mcpServers": {
    "arke": {
      "command": "/absolute/path/to/arke/.venv/bin/arke-mcp"
    }
  }
}

Restart Claude Desktop.

Claude Code

claude mcp add arke /absolute/path/to/arke/.venv/bin/arke-mcp

Cursor

Settings → MCP → Add new MCP. Type: stdio. Command: /absolute/path/to/arke/.venv/bin/arke-mcp.

4. Verify

In your AI assistant chat:

Use the arke search tool to find authorities on "abuse of dominance excessive pricing".

You should see top-K results with neutral citations and snippets.

For a full stress-test:

Use the arke stress tool on this draft: [paste a paragraph from your skeleton].

Expect 30-60 seconds. The reply is markdown with ## headed limbs and > blockquoted verbatim citations.

Configuration

scripts/install.sh writes a working .env for the OpenAI path. Override any of these to point at a different inference backend (e.g. a local llama-cpp-server):

• EMBED_URL, FAST_URL, STRONG_URL — base URLs. Default: https://api.openai.com. Each must speak /v1/embeddings (embed) or /v1/chat/completions (fast/strong).
• EMBED_MODEL, FAST_MODEL, STRONG_MODEL — model names. Required.
• INFERENCE_KEY — bearer token. Required (the server fails fast if empty).
• ARKE_HOME — state directory. Default: ~/arke/. Holds the corpus, indexes, mailbox, and logs.

The same arke/inference/client.py HTTP code path is used regardless of backend; the network boundary is the actual no-egress guarantor, not config.