aksika

Appearance

abTARS ​

Meet the latest kid on the block. An agentic framework designed to survive — and do the job.

abTARS is an autonomous AI agent framework that runs 24/7 on your hardware. It connects frontier models to chat platforms, manages persistent memory, executes tasks on schedule, self-heals from failures, and coordinates work across a distributed swarm of peer instances. Zero cloud dependency. Your rules.

Design Philosophy: CIA Triad ​

Every feature in abTARS maps to one of three pillars:

Confidentiality — Your data stays classified. NATO-style memory tiers, AES-256 encryption at rest, secrets vault with strict permissions, Ed25519 signed peer communications, injection scanning on untrusted payloads.

Integrity — The system never lies about its state. Peer trust levels gate execution depth. Memory contradictions auto-resolve (old facts expire). Atomic writes prevent corruption. Doctor validates everything continuously.

Availability — It runs, it recovers, it never stays down. Three-legged supervision (watchdog → bridge, OS → watchdog, circuit breaker → rollback). Bad deploys auto-rollback in ~30s. Corrupt state self-heals. Stress-tested: kill any component, it comes back.

→ Resilience & Stress Tests

Plus: Distributed Agent Swarm ​

abTARS instances find each other via gossip, delegate work based on capabilities and load, transfer artifacts, and deliver results via callbacks. One agent is useful. A swarm is unstoppable.

→ Agent Swarm

Plus: Pi Integration ​

abTARS integrates Pi as a symbiotic peer — Pi's provider engine powers Direct API with ~36 providers and prompt caching, Pi's TUI gives you a terminal face, and Pi's coding agent runs supervised coding tasks. Additive, reversible, zero coupling.

→ Pi Integration

Features ​

  • Multi-platform — Telegram, Discord, IRC (more coming)
  • Multi-provider — ollama, OpenRouter, Kiro CLI, Gemini CLI, Codex
  • Automatic fallback — leaky-bucket health tracking, progressive backoff
  • Persistent memory — powered by abmind (in-process, multi-layer recall)
  • Secrets vault — AES-256-GCM encrypted at rest, passphrase-derived key
  • Scheduled tasks — cron-style with retry, logging, notifications
  • Peer-to-peer — A2A protocol over Tailscale (JWT + digital signatures)
  • Skills — self-organizing procedural knowledge the agent learns during sleep
  • 3-legged supervision — auto-recovery from crashes, bad deploys, and corruption
  • Auto-rollback — broken code detected in 30s, previous version restored

→ Installation guide

Community ​