Open source · self-hosted · v0.3.0

A self-hosted AI agent that argues with itself before it answers you.

Execution, quality review and final sign-off are handled by models from different vendors. Swap between 13 providers. Reach it from 7 messaging platforms. Everything it remembers stays on your own machine.

Get started Source code
SuperCrzy Bot
Introduce yourself.
Thought for a moment · 3 tools used

I'm SuperCrzy Bot — an AI agent that runs on your own machine. You bring the model keys, I keep the work.

Content ecosystem
397 in-house stories, retrieved with exact publication dates
Multi-model senate
execution, QA and review sit on different vendors' models
Guardrails
three-verdict approval, then an OS-level sandbox
Reach
7 messaging platforms plus web and desktop
news_searchcollab_runfile_readweb_searchdeep_researchmemory_recall+38
Where does my data live?
file_read · ~/.supercrzy
supercrzy.db — sessions, messages and memory, in one SQLite file
data/skills — Agent Skills standard, portable to other agents
data/checkpoints — every stage written as it completes, resumable
Describe what you want to build — or type / and @ files⌘↵
397In-house stories
13Model providers
7Messaging platforms
43Built-in tools

ARCHITECTURE

Three packages, one process, your disk.

A pnpm monorepo. The server is the only brain — every interface talks to it, and it reads and writes a SQLite file in your home directory. Nothing is routed through our infrastructure.

Interfaceshow you reach it
Web UIVue 3 · 15 views DesktopElectron CLIterminal 7 chat adaptersWeCom · DingTalk · Feishu · QQ · Telegram · Slack · Discord
HTTP + SSE
CoreKoa · one process
Agent loopturn / tool / stream Tool registry43 tools · 2-tier loading API routes26 modules Memory + skillsFTS5 retrieval Permission + OS sandbox3 verdicts · seatbelt Collaboration engineexecutor / QA / senate
reads + writes
Your diskno cloud account
~/.supercrzy/supercrzy.dbSQLite · 25 tables data/skillsAgent Skills standard data/checkpointsresumable runs

One outbound dependency by design: the content feed at supercrzy.com. Models, tools, memory and history are all yours to bring or keep. MIT licensed.

THE SENATE MODEL

One model will miss things. Several models from different vendors, checking each other, usually won't.

A collaborative task is not "finish it and hand it over." The deliverable must clear QA first, then face a vote — and the executor, the reviewer and the senators are different models. None of them rubber-stamps another's work.

Execution
DeepSeek

Produces the deliverable stage by stage. Every stage is written to the database as it completes.

Quality review
Kimi

Independent scoring. Below the threshold, the work is sent back to execution.

Senate
DeepSeek + Kimi

"Functional chair" checks completeness; "logic debater" checks internal consistency. Each senator votes independently.

Verdict rule: a majority in favour, with no high-severity issue. Otherwise the task is automatically returned for rework — up to 2 rounds. A senator that is unavailable abstains rather than blocking the whole review.

Rejected means reworked, not re-prompted

If QA or the senate rejects the result, the task loops back to execution on its own. You don't have to ask again.

  • Convergence capped at 2 rounds so it can't spin forever
  • Configurable quality threshold, default 70

Two gears: supervised or autonomous

Same engine, two modes. Semi mode stops at every stage for your decision; senate mode runs the whole pipeline by itself.

  • Supervised runs can be escalated to a senate review in one click
  • Stage timeouts with cause reporting (timeout / quota / unreachable — separated)

CONTENT ECOSYSTEM

Not "it can search the web." It ships with a content ecosystem.

A generic agent scrapes the open web and then guesses publication dates. SuperCrzy Bot reads our own feed — 397 stories, updated daily, each with an exact publication date. Ask what happened today and it won't get the date wrong.

SuperCrzy Bot — the in-app AI news board, reading the in-house feed
The news board inside the app. Every card comes from our own feed — which is why the dates hold up.

It doesn't flow one way. When the agent answers, it sends the reader back to the full story; when people land on the publication, they can pick up the agent. Each side is the other's front door.

MODEL FREEDOM

Thirteen providers. Swap them whenever you want.

Bring your own keys — including a model that runs entirely on your own machine. Nothing is bundled with the app and nothing is proxied through us.

SuperCrzy Bot — model management listing thirteen API providers
Provider list inside the app: local Ollama, NVIDIA NIM, Qwen, DeepSeek, OpenAI, OpenRouter, Anthropic, xAI Grok, Google Gemini, Mistral, Zhipu GLM, Kimi and MiMo.

SAFETY

You can let it touch your files because there are two gates, not one.

Most agents give you "allow" or "deny". SuperCrzy Bot splits a shell command into segments and judges each one, then puts an OS-level sandbox underneath.

Three verdicts, not two

Every tool call is classified as allow, ask or block. Tools that require explicit interception are checked before the allowlist — being on an allowlist can never silently skip a block.

  • Shell commands are judged per segment; quotes and heredocs are understood, a single & doesn't split
  • Destructive operations are graded: out-of-scope, recursive or privileged → blocked outright

An OS sandbox underneath

Tool execution runs inside an operating-system sandbox. Readable and writable paths are allowlisted — if the approval gate is bypassed, this one still holds.

  • Three permission levels: read-only / ask before acting / full access
  • Paths outside the working directory and project root are denied by default

AI BROWSER

It can drive a real browser — and you can watch it work.

Point it at a URL and it browses like a person: clicks, selects, types, scrolls. The live view sits in the side panel, so you see the page it is reading while it reads it. You can take over at any moment — it is the same page.

Screen recording from the app: the agent opens supercrzy.com and scrolls through it. Muted, loops automatically.

MULTI-PLATFORM

One memory. Seven platforms can reach it.

Sessions, memory and the skill library are shared across every endpoint. Ask it something from WeChat, then pick the thread back up on your desktop.

China

  • WeCom
  • DingTalk
  • Feishu
  • QQ

Global

  • Telegram
  • Slack
  • Discord

First-party

  • Web (any browser)
  • Desktop (macOS)
  • CLI

QUICK START

Two commands. No cloud account, no trial credit, your own keys.

Requires Node.js 22. On Windows use start.bat. First launch walks you through entering your own model key — 13 providers supported, including a fully local Ollama.

$git clone https://github.com/SuperCrzy/Supercrzy-Bot.git
$./start.sh