docs
how to talk to zeus + what zeus can do
about zeus
Zeus is a personal AI chief of staff. One Telegram bot, one Python process, one SQLite file. ~30 deputies, 260+ tools.
You talk to Zeus on Telegram (text, voice notes, photos). Zeus calls tools to do work — wallets, swaps, research, content, code reviews, self-modification. Every destructive action waits for an explicit yes/no before it fires.
architecture
Telegram bot ─┐
│
HTTP webhooks (Helius/Alchemy) ─→ │ ┌─ confirmation gate
▼ │
ZeusBrain (Sonnet → Haiku → Opus
│ fallback chain)
│
┌───────┴───────┐
│ ~30 deputies │ each exposes tools
│ ~260 tools │ the brain calls
└───────┬───────┘
│
┌───────────┬───────┼────────┬────────────┐
▼ ▼ ▼ ▼ ▼
wallets trading memory markets self-devcore ideas
- Brain → deputies → tools. One Claude reasoning loop. Deputies are Python classes; each exposes tools the brain calls directly. No agent-to-agent chatter.
- Confirmation gate. Sends, posts, swaps, code edits, key storage — all surface a one-line summary and wait for yes / no / "I accept the risk". Some tools are mode-aware: paper bypasses, live triggers.
- Self-modifying. When you ask for something Zeus doesn't have, the dev brain researches the API, generates code, sandboxes it (Docker), iterates on test failures, proposes the diff, applies it, restarts. Triple-layer rollback (git → FS snapshot → DB diff).
- Vault. Private keys + API tokens are Fernet-encrypted at rest, key file lives outside the repo.
- Multi-model fallback. Sonnet → Haiku → Opus on rate-limit / credit / 5xx. Daily cost cap short-circuits before any API spend.
what this is not
- Not a multi-tenant SaaS. Single-user by design.
- Not audited. Treat the wallet code, sandbox, and gate as trust-but-verify.
- Not financial advice. Trading deputies default to paper.
- Not stable. The codebase rewrites itself; expect drift.