TARS
A self-hosted AI agent with a visible, auditable brain. Every decision is traceable through a hash-chained action ledger. Every personality trait is a Bayesian probability, not a hardcoded prompt. Every model call comes with a cost receipt.
By the Numbers
- 6-10K — Lines of Code
- 37 — Test Files
- 9 — Implementation Phases
- 100% — Auditable
The Problem
AI agents are black boxes. You can't see why they made a decision, you can't verify their reasoning, and you can't tune their personality without rewriting prompts. Most agent frameworks optimize for capability at the expense of transparency.
TARS takes the opposite approach: every action is logged in an immutable ledger, every personality trait is a tunable probability distribution, and every model call is tracked with cost receipts.
Key Features
- Bayesian Genome System — Personality traits are probability distributions, not static strings. The agent's behavior evolves based on interaction feedback — measurably and predictably.
- Hash-Chained Action Ledger — Every action, decision, and model call is logged in a tamper-evident chain. Full audit trail from input to output. Immutable history.
- Multi-Tier Model Router — Intelligent routing between model tiers based on task complexity. Simple queries go to fast/cheap models, complex reasoning goes to capable ones. Cost receipts on every call.
- Self-Hosted — Runs entirely on your infrastructure. No data leaves your network unless you route a model call externally. Privacy by architecture, not policy.
- 37 Test Files — Comprehensive test coverage across all 9 implementation phases. Each system component is independently testable and verified.
- Phased Architecture — Built in 9 deliberate phases — from core runtime to memory to model routing to audit. Each phase is stable before the next begins.
Tech Stack
Python · SQLite · Pydantic · pytest · hashlib · NumPy