AUTONOMOUS AI ORCHESTRATION

Run Your Own Model.
Zero Censorship.
Zero Platform Tax.

A self-hosted, uncensored AI agent stack. Drive an AI agent from WhatsApp through an MCP tool server, and serve an uncensored model on your own GPU with the bundled llama.cpp image. You run every piece; you choose the model endpoint.

Get Started View on GitHub →
0 MCP Tools
$0 Platform Fee
Zero Content Filters
100% Self-Hosted

We build infrastructure, not guardrails.

Morality is a human prerogative — not a system configuration.

01

User Sovereignty

OpenTard doesn't filter, rewrite, or refuse. Your prompt, your model, your responsibility. We pass intent without editorial interference.

02

Your Prompt, Not Ours

The agent's system prompt is a config value you own and can read in full. No vendor filtering layer, no output rewriting, no behavioural patch you didn't write.

03

No Middleman

The agent, message bridge and storage all run on your machine, and the model endpoint is whichever one you configure — the Claude API by default, or your own llama.cpp server. There is no OpenTard service in the path. We can't leak what we never receive.

Why OpenTard

$

Lowest Cost

Serve the model yourself on hardware you already own. No markup, no platform tax, no subscription, no per-token bill at all when you run locally.

  • llama.cpp server with the model baked into the image
  • No monthly fees, minimums, or seat counts
  • OpenAI-compatible API — point existing clients straight at it
  • Or plug in a hosted API when you want a frontier model
◈

Yours End to End

Self-hosted from the model up. Services bind to loopback by default, the WhatsApp session lives in a volume you control, and prompts never transit a third party you didn't configure.

  • Runs entirely in your own Docker stack
  • Agent stack binds to loopback by default
  • Sender allowlist gates who can drive the agent
  • Shell tool ships disabled — opt in deliberately
⊘

Zero Restrictions

No content policy is enforced by this stack. Run an uncensored model and you get it exactly as its weights behave — no filtering, rewriting, or refusal logic bolted on top.

  • No filtering or rewriting layer
  • System prompt is yours to set or clear
  • Default image ships an uncensored Qwen3.5 build
  • Only your model's own behaviour applies — not our policy

Architecture

The agent stack runs on your machine and calls the Claude API. The llama.cpp server is a separate, self-hosted endpoint you can run alongside it for fully local inference.

Inference
You
→
tard agent
→
Claude API
OpenAI-compatible client
→
llama.cpp
→
Your GPU
your machine
Messaging
Agent
→
ot-wap (MCP)
→
WhatsApp
! 15 MCP tools expose WhatsApp to the agent — send, receive, groups, media. The session is linked once by QR and stored in a volume you own.

Get Started

Clone the repos side by side, set a key and a webhook secret, then bring the stack up with Docker.

terminal
$ git clone https://github.com/opentard/apex && git clone https://github.com/opentard/wap
$ echo "ANTHROPIC_API_KEY=sk-ant-api..." >> apex/.env
$ echo "WAP_WEBHOOK_TOKEN=$(openssl rand -hex 32)" >> apex/.env
$ cd apex && docker compose up --build
 
● wa-bridge listening (internal)
● ot-wap MCP server on 127.0.0.1:8080
● tard dashboard on 127.0.0.1:5000
Scan the QR from the GetWhatsAppQrCode tool to link.
Ready.
View on GitHub Documentation →