anz33

One local endpoint for every AI engine. — anz33.com

anz33 is a small daemon that sits in front of every inference engine on (and around) your machine — Ollama, LM Studio, llama.cpp, and optional cloud endpoints — and exposes a single OpenAI-compatible API. Apps point at http://127.0.0.1:3300/v1 and anz33 routes each request to the best engine, using measured tokens-per-second on your hardware and your current power state.

Why: laptops now ship five incompatible AI stacks and most apps hardcode one. The router is the missing layer — pick the engine per request, not per app.

Quick start

uv tool install anz33        # or: pip install anz33
anz33 serve                  # http://127.0.0.1:3300/v1

Point any OpenAI-compatible client at it:

curl http://127.0.0.1:3300/v1/chat/completions \
  -d '{"model": "auto", "messages": [{"role": "user", "content": "hello"}]}'

"model": "auto" lets anz33 choose. Name a specific model and anz33 finds whichever engine hosts it.

Commands

Privacy (L1)

Requests are evaluated before routing: a matching local_only rule removes every cloud engine from consideration, deny blocks the request outright. When a request is allowed out, the PII airlock replaces emails, phone numbers, cards, IPs and known names with placeholders and restores them in the response — the cloud sees shapes, not people. Every decision appends to a hash-chained audit log; editing or deleting any record breaks the chain and anz33 audit says so.

Memory vault (L2)

Memories are plain markdown in ~/.anz33/memory/. The router retrieves the relevant few and injects them into any model, so a local 8B remembers you as well as a frontier cloud model. anz33 also captures durable facts from your own turns automatically, locally.

Skills, chat, household (L3, L5)

Chat UI at /anz33chat; dashboard at /anz33dashboard. Select a skill per request:

{"model": "auto", "skill": "fraud-check", "messages": [...]}

Built-ins: homework (explains, never writes the essay), fraud-check, translate, summarize, private-notes, explain-simply. Sensitive skills are local-only by construction.

Routing policy

1. Exact match, local first — the requested model on the fastest engine that has it (by your anz33 bench measurements).
2. Battery-aware — on battery below 50%, big models (>15B) are swapped for the smallest capable local model. Disable with "battery_saver": false.
3. Cloud is opt-in — remote endpoints are only used if "allow_cloud": true in ~/.anz33/config.json. API keys are read from environment variables, never stored.
4. auto — fastest measured local model; falls back to a size heuristic until you run anz33 bench.

On Apple silicon, anz33 also probes an mlx-lm server for native-speed inference: pip install mlx-lm && mlx_lm.server --port 8081.

Every response carries anz33-engine and anz33-route headers explaining the decision.

Configuration

~/.anz33/config.json (optional — zero config works):

{
  "port": 3300,
  "allow_cloud": false,
  "battery_saver": true,
  "local_engines": {
    "ollama": "http://127.0.0.1:11434",
    "lmstudio": "http://127.0.0.1:1234",
    "llamacpp": "http://127.0.0.1:8080",
    "mlx": "http://127.0.0.1:8081"
  },
  "remotes": [
    {
      "name": "anthropic",
      "base_url": "https://api.anthropic.com/v1",
      "api_key_env": "ANTHROPIC_API_KEY",
      "models": ["claude-sonnet-5"]
    }
  ]
}

Development

uv sync --extra dev
uv run pytest
uv run anz33 serve

Pricing

anz33 is subscription software: $20/month, everything included — routing, benchmarking, dashboard, remotes, priority support. New installs get a 14-day evaluation trial, after which the daemon requires an active subscription (anz33 activate <key>).

Proprietary — © 2026 AZMX. All rights reserved.