PublishedFramework tierShowcaseFree for individuals
Black Unicorn Command Centre
BUCC is the production multi-agent operations platform we built to run our own 25-agent fleet. Not a product we sell. The open hood on how we engineer a secure agentic framework.
Overview
Most multi-agent deployments fail because the infrastructure isn't mature. Teams stitch together frameworks, APIs, and custom code without thinking about governance, audit, memory, capacity, or human oversight. They bolt those layers on later. That's when things get expensive and fragile.
We asked ourselves a different question: what would it actually take to run 25 AI agents in production, with governance, memory, and security as first-class citizens from day one? BUCC is the answer.
The Black Unicorn Command Centre is a full-stack, self-hosted multi-agent operations platform. Not a framework wrapper. Not a research project. A production system that has been running our internal fleet for 18 months. We are showing it publicly because the community needs to see what production multi-agent operations actually look like, and because BUCC is the reference implementation behind our Agentic Framework Design & Operations service.
The receipts:
- 390+ API endpoints across platform, governance, content, and lifecycle domains
- 130 relational tables tracking agent state, work assignments, governance decisions, approvals, audit trails, and financials
- 25-agent fleet across finance, cybersecurity, research, ops, comms, and planning
- 3-layer LLM routing distributing inference across L1 local Ollama, L2 subscription, and L3 pay-per-token providers
- 3 approval tiers: T1 auto-execute, T2 notify, T3 block-until-approved
- 5 circuit breakers (CB-1 through CB-5) for graduated governance control, from soft pause to hard stop
- 3-tier persistent memory via mem0: Tier 1 global, Tier 2 agent-specific, Tier 3 session
- 9-widget CEO Dashboard surfacing fleet health, approvals, security, financials, and ops in real time
Governance-first design. Governance isn't a feature you add after shipping. It's an architecture question. CB-1 through CB-5 fire on policy violations before actions reach the world. The T1/T2/T3 classification makes human oversight proportional to risk instead of either rubber-stamping everything or blocking everything.
Fail-closed security. The Data Sanitization Proxy classifies and routes outbound LLM calls. Sensitive contexts route to L1-only (DSP-high). Every decision, action, approval, and error is written to a tamper-evident audit trail. When the system can't decide, it stops.
Memory as infrastructure. Agents in isolation are amnesiacs running in parallel. Real fleets remember. The 3-tier memory lets agents learn from past work, coordinate on shared goals, and stop restarting cold on every task. Tier 1 holds global knowledge, Tier 2 holds agent-specific learning, Tier 3 holds session context.
Human oversight by design. The CEO Dashboard, the approvals queue, and the Ops Center are not bolted-on admin pages. They are the operating surface. The fleet is steered, not just observed.
The stack: Next.js operator UI, FastAPI control plane, PostgreSQL state, mem0 memory layer, OpenClaw agent runtime, Docker Compose deployment. 26 operator pages, 47 FastAPI routers, 88 backend services, 70+ skill documents across the fleet.
This isn't theoretical. It's running right now.
Tech Stack
Next.jsFastAPIPostgreSQLmem0OpenClawDocker
Key Features
25-agent fleet across finance, cybersecurity, research, ops, comms, and planning
390+ API endpoints across platform, governance, content, and lifecycle domains
130 relational tables for agent state, work, governance, approvals, audit, and financials
3-layer LLM routing: L1 local Ollama, L2 subscription, L3 pay-per-token
3 approval tiers: T1 auto-execute, T2 notify, T3 block-until-approved
Circuit breakers CB-1 through CB-5 for graduated governance, from soft pause to hard stop
3-tier persistent memory via mem0: Tier 1 global, Tier 2 agent-specific, Tier 3 session
9-widget CEO Dashboard for real-time fleet health, approvals, security, and ops
Tamper-evident audit trail of every decision, action, approval, and error
Data Sanitization Proxy (DSP) for outbound LLM call classification and routing
Approvals queue with human-in-the-loop, YOLO mode, and analytics
OpenClaw agent runtime with 70+ skill documents across the fleet
Screenshots
Powering Our Services
Black Unicorn Command Centre is used directly in the following services:
LLM & AI Security
Comprehensive security assessments for LLM-powered applications. From prompt injection testing and AI agent security to multi-agent operations and custom model hardening.
Agentic Framework Design & Operations
Design, build, and secure multi-agent AI systems. From architecture to deployment, with battle-tested orchestration patterns.
A demo version of the Black Unicorn Command Centre is available at: bucc.blackunicorn.tech
Agentic implementation and consulting engagements are reserved to existing Black Unicorn customers. Join the waiting list to be notified when new slots open.