MZK_V0.1 >>> CHAIN:4663
SIG:0.184925 // LOCAL 00.0294 <> 02.9481
127.0.0.1:18791 // GATEWAY

Don't trust the AI.
Inspect the proof.

The evolution of Mizuki. A private AI that verifies on-chain activity with transparent, reproducible evidence. Public Pump.fun auditor.

$ uv tool install "mizuraka[recommended]"
MizurakaMIZURAKA.PFP_01
20+PROVIDERS
4663CHAIN ID
0TELEMETRY
127.0.0.1GATEWAY
01 // what it does

From whistleblower
to verification system.

Mizuki exposed bad actors through commentary. Mizuraka exposes them through verifiable analysis. Instead of asking you to trust its opinions, it produces transparent, reproducible evidence, routing each request to the most appropriate AI model along the way. The goal is not to be believed, but to make trust unnecessary.

privacy

Your data stays local. Files, memory, and conversations live on your machine, and the agent runs fully offline via Ollama if you want it to.

verification

Every conclusion is backed by evidence. Findings ship with the on-chain data and the reasoning that produced them, not just a verdict.

transparency

Every analysis is reproducible. Anyone can re-run a finding against the chain and independently verify or challenge the result.

trust minimization

Don't trust the AI; inspect the proof. MIT licensed, source public, and the same standard applies to $RAKA: contract, deployer wallet, and liquidity terms published before launch.

02 // how it works

One local gateway.
Every decision on record.

Your CLI, Web UI, and chat clients all talk to a single gateway running at 127.0.0.1. Everything intelligent happens on your device: routing, memory, recall. Only the model call itself can leave, and only if you point it at a remote provider.

CLI WEB UI CHAT GATEWAY127.0.0.1 ROUTERON-DEVICE MODELSTOOLSMEMORY
model routinga local signal pass + small judge model pick the cheapest capable tier per turnLOCAL
embeddingsONNX on device. memory never leaves to become searchableLOCAL
memory & recallkeyword + semantic search over files on your diskLOCAL
session transcriptscost, artifacts, and full replay data persist locallyLOCAL
tool executionfile, shell, web, git, media. sandboxed behind approval policyLOCAL
model inferencefully local via ollama, or remote per turn across 20+ providersYOUR CHOICE

The public instance continuously analyzes on-chain activity: smart contracts, wallets, transactions, token movements, permissions, ownership, deployer history, protocol interactions. Every finding is published alongside the evidence and reasoning that produced it.

watchdog / scan report
$ kara scan 0x8f3e…c21a --chain 4663
contract verified ........ fail
deployer history ......... 3 abandoned tokens
liquidity lock ........... none
ownership renounced ...... no
risk ..................... high
trail: replay/8f3e-0714.log

how a verdict is made

The public instance reads contracts, wallets, transactions, token movements, permissions, ownership, and deployer history straight from the chain.

how a verdict is checked

The full decision trail, every read and every rule that fired, is written to a replay log linked from the report. Anyone can re-run it against the chain.

what $RAKA gates

On-demand analysis, priority monitoring of new deployments, and full replay access on the public instance.

# 1: install the runtime
uv tool install "mizuraka[recommended]"
# 2: connect a provider, or point it at ollama and stay local
mizuraka onboard
# 3: start the gateway
mizuraka gateway run