Arcana lets you configure and run AI agents as tarot card archetypes: typed personalities where system prompt, temperature, and tone are compiled automatically from the card.
Free forever. Apache 2.0 licensed. Python 3.11+.
No consistent character, no overview, nobody in charge. You can't tell what they know, why they behaved the way they did, or who to hold responsible when they go off-track.
Agents are a model name and a system prompt. No archetype, no voice, nothing that makes one reliably different from the next. Same tool, different answer, no explanation why.
Personality evaporates between prompts. The same agent runs differently every session because there's nothing anchoring who it is, just a system prompt you rewrite by hand.
Nothing logs intent. You see the output, not the reasoning: the cost, not the cause. When something breaks, you're debugging a black box.
Arcana is built on one premise:
agents should have archetypes, not just prompts.
Each of the 22 Major Arcana is a complete behavioral archetype: enough structure to make your agent reliably itself, session after session.
One card sets the archetype. Add modifier cards and the engine blends them into a single compiled config, temperature-adjusted, and checked for tension before it ever runs.
The primary card defines the agent's full archetype. Modifier cards shift the temperature at the weights shown and add a targeted tone layer. Compiled into one system prompt, not assembled at runtime.
The engine reads your combination like a spread. Archetypes that amplify each other get a synergy flag; conflicting ones get a warning before you ship.
The modifier recomputes temperature, not just tone. Hermit (0.35) with a touch of Sun (0.75) runs measured-but-brighter. Visible, deterministic, overridable.
Local or hosted, your choice. Your keys never leave your machine.
Browse the 22 Major Arcana. Each carries a defined archetype, temperature, and tone.
Name it and pick its card. Arcana compiles the config: system prompt, temperature, tone.
From the reckless Fool to the all-seeing World: not a list of features. A cast of characters.
Common starting points: Hermit (IX): research agent · Empress (III): writing agent · Magician (I): automation agent · Justice (XI): code review
Ollama, Anthropic, or any OpenAI-compatible endpoint. Local models work out of the box.
Three tiers (private, shared pools, global) so agents keep what's theirs and pool what's shared. Shipping soon.
A meta-agent that routes tasks, audits for reversed cards (agents running outside their archetype), and writes your daily briefing. The one card that cannot be reversed. Shipping soon.
Start local in one command. The deck is waiting.