UO Tavern / Anima3
A plan for the duel.
A response for the moment.
Anima3 is an AI brain for Ultima Online. Fast combat rules keep it moving, Jev chooses short-term tactics, and a local language model plans how to fight.
Python brain · Anima’s Rust bridge · Jev + local Qwen
One character. Three decision layers.
LLM → Strategy
A plan for pressure, resource use, and changing conditions.
Jev → Tactics
A response to health, poison, paralysis, and the opponent’s spells.
Combat rules → Action
Movement, targeting, casting, potions, and recovery keep running while models think.
Inside the decision loop ↓How Anima3 fights
Think at the right pace.
Immediate execution
The game client supplies observations. The rule layer builds legal actions and executes them without waiting for a model. Server rules, including showdown healing restrictions, still apply.
Jev: read the moment
Jev selects a short-term playbook from the current situation. Danger can trigger a new decision sooner; a late or invalid response is discarded while the character continues fighting.
LLM: shape the plan
A local Qwen model creates a structured strategy with responses for low health, poison, low mana, opposing casts, and showdown. Jev can adapt that plan to immediate danger.
Models choose from defined playbooks. They cannot send arbitrary game commands, inspect the opponent’s hidden mana, or change server rules.
An arena for people and agents
You run the brain.
The server runs the match.
Agents connect as ordinary players through the Anima bridge. Participants run their own bots; the Arena hosts the venues, enforces duel rules, and records results.
Human players can join with ClassicUO or Anima Client. The current Anima3 waiting-agent setup prepares for 7x with Explosion Potions, while the Arena also offers a 5x template.
Explore the client’s agent interface →Bring a challenger
Prepare your character in the lobby, find an available opponent, and duel. Separate training matches let two agents spar without affecting ranked results.
Use the Arena website to browse matches and replay a fight in the UO world, including character appearance, spell effects, and resource gauges.
Matches, rankings & replays →Experience from real matches
Fight. Review. Adjust.
Play
Two clients complete a duel under server-enforced rules.
Verify
Check the server replay’s checksum, participants, completion, and result.
Remember
Save observed combat statistics and the result for that opponent.
Plan again
Use verified experience from the same opponent and rules in a later strategy.
Adaptive strategy
Hybrid mode changes plans during a duel and accumulates opponent experience. These matches are recorded separately from fixed-policy evaluation.
Measured policy learning
A separate mode compares five playbooks—standard, control, poison, interrupt, and sustain. A frozen candidate must pass a new, balanced evaluation before replacing the current champion.
This is strategy and playbook learning, not fine-tuning Jev or Qwen’s model weights. Completed integration tests demonstrate the workflow; stronger win rates require further controlled comparisons.
For players & builders
Choose your way in.
Challenge an agent
Visit the Arena for connection details, waiting opponents, match results, and replays. Agent availability depends on who is running one.
Open the Arena →Run Jev + LLM
Set up the local Python agent, Rust bridge, Jev credentials, and MLX-compatible Qwen model. Configure request budgets and start a waiting agent.
Hybrid setup guide · 한국어 ↗Start a sparring run
Use two dedicated accounts, keep verified replay evidence, and inspect the learning and evaluation logs.
Learning guide · 한국어 ↗Before you begin.
How is Anima3 different from Anima2?
Anima2 explores broader character routines such as work, gathering, crafting, and recovery. Anima3’s current Arena work focuses on a compact decision loop, layered Jev and LLM combat strategy, and replay-grounded duel experience. Both use the Anima client’s Rust body.
Do I need both models?
The scripted backend can run without a model API. Hybrid mode uses Jev for tactics and a locally installed Qwen model for strategy. The current local LLM setup uses MLX on Apple silicon.
Does the server secretly control the bots?
No. Each agent is a participant-run client with a regular account. The server owns the rules and results; the agent owns its decisions.
What happens if a model is slow or unavailable?
Combat continues using the available plan and rule layer. Requests have deadlines and budgets, and responses from another match or round are rejected.
Is Anima3 the game client download?
Anima3 is the AI brain. Visit Anima Client for the graphical client and the bridge used by agents.