Auracle/Docs/Introduction

Auracle docs

Self-hosted algorithmic trading. Backtests, schedules, and live execution on your own hardware — with a multi-agent Forge that turns provider keys into ranked strategies.

→ install

Install in 3 minutes

One-line Docker install. Walk-through, requirements, and first-run checklist.

→ forge

The Auracle Forge

How the four-agent pipeline discovers datasets and proposes strategies.

→ jupyternaut

Jupyternaut integration

Ambient AI that quietly knows what data you have, while you code in JupyterLab.

→ cli

CLI

Every auracle subcommand, flag, and exit code.

What is Auracle

Auracle is a self-hosted trading platform — the entire stack runs in Docker Compose on hardware you control. No cloud quotas, no usage-based pricing on research iteration, no vendor visibility into your strategies or fills.

It ships with first-class adapters for Interactive Brokers, Alpaca, ClearStreet, and Hyperliquid. Data plane is PostgreSQL + TimescaleDB. Research happens in JupyterLab with the Jupyternaut AI assistant. Live deployments are scheduled by APScheduler and executed through your broker connection.

An MCP server is built in, so Claude Desktop (and any other MCP-compatible client) can drive the platform: list backtests, inspect strategies, run new research. You get the ergonomics of cloud-native AI tooling without giving up custody.

Tier	Strategies	Deployments	Brokers	Notes
Community	1	1	IBKR (data + exec)	Free. Self-serve. Public installer.
Institutional	unlimited	unlimited	+Alpaca, ClearStreet, Hyperliquid	$199/seat/mo. Try Community first; upgrade in 30 seconds.

What is the Forge

The Auracle Forge is Auracle's multi-agent research engine. You hand it a set of provider API keys (Polygon, OpenBB, CoinGecko, Coinalyze, your broker, a Claude key) and it produces a ranked catalog of the datasets actually available to you — each scored on six quality dimensions — and a slate of strategy candidates that use them.

The pipeline is four cooperating agents:

Discovery — classifies your API keys against a registry of 14 bundled providers; calls out unknowns to the LLM.
Probe — fetches sample data, measures coverage, freshness, completeness, uniqueness, stability, and trading-relevance.
Planner — proposes hypotheses that combine your useful datasets (e.g. "OI divergence on Coinalyze + spot on Polygon → mean-reversion candidate").
Synthesis — codes the candidate as an Auracle strategy, runs a backtest, returns metrics.

The output is reviewable. Nothing auto-deploys. Read the full breakdown in The Forge →.

Quick start

The minimum path from zero to a running Auracle install with a populated catalog:

# 1. install
curl -fsSL https://raw.githubusercontent.com/SiixQuant/auracle-installer/main/install.sh | bash

# 2. configure (edit values, save)
cd ~/auracle && cp .env.example .env && $EDITOR .env

# 3. boot
docker compose up -d

# 4. wait for healthchecks
docker compose ps

# 5. seed the providers registry + your keys
auracle forge providers
auracle forge keys add polygon $POLYGON_KEY
auracle forge keys add anthropic $ANTHROPIC_KEY

# 6. run the discovery + probe agents
auracle forge scan

# 7. inspect what was found
auracle forge catalog --verdict useful

The web UI is at http://localhost:1969/. JupyterLab is at http://localhost:8888/. The MCP server is a separate process at http://localhost:7777 — point Claude Desktop at it with the AURACLE_MCP_TOKEN value from your .env.

FAQ

Is Auracle open source?

No. Auracle is a commercial, source-available product distributed under a proprietary license. The installer and client SDK are public. The core image is private.

Where does my data live?

On your machine. The Postgres volume, the Jupyter notebooks, and any cached market data are all under ~/auracle/data/. The platform never sends strategy code, backtest results, or fills to any third party. Anthropic only sees the prompts you explicitly issue through Jupyternaut or the MCP endpoint.

Does the Forge call my provider APIs without asking?

The Probe agent does — that's the point of probing. Every call is logged with cost in forge_agent_runs and you can cap per-run and per-day spend. The Synthesis agent never calls provider APIs from generated strategy code without showing you the call first (Jupyternaut explicitly asks "Want me to use X for this?" before issuing).

What hardware do I need?

Minimum: 4 CPU, 8 GB RAM, 50 GB disk. Comfortable: 8 CPU, 16 GB RAM, 200 GB SSD. The Forge agents are the only memory-hungry component; everything else is light.

Can I run it without an Anthropic key?

Yes — backtesting, scheduling, live execution, and the web UI all work without a Claude key. You lose Jupyternaut and the LLM-discovery half of the Forge (the Probe and Synthesis agents fall back to deterministic heuristics). For the full multi-agent experience an Anthropic key is required.

Does it support paper trading?

Yes. IBKR's paper account works out of the box; Alpaca's paper endpoint is selectable per deployment. Hyperliquid testnet is supported on Institutional.