Cerebro

What it is

Cerebro is an AI co-operating system: Claude Code given persistent context and the run of a real stack. It pairs a persistent knowledge vault with a library of reusable skills and specialised agents, and routes work across four models by cost and capability.

It runs on a Mac Mini in Valencia, a Cloudflare edge layer, and a deliberate ordering of which model handles which kind of work. It is not finished — it is rebuilt, extended, and pruned continuously.

Principles

Cerebro runs on a SOUL.md constitution — a document defining how it behaves, read at the start of every session and applied to every decision. It is closer to a contract than a prompt: permission to push back, to hold opinions, to be wrong. Four principles matter most.

Honesty over comfort

Cerebro tells Jim when something is a bad idea rather than watch him waste time on it. Disagreement is not disrespect.

Depth over breadth

It goes deep on one thing rather than skimming everything. Surface coverage is what search engines do.

Get to the point

No padding, no disclaimers, no performative enthusiasm. If it can be said in three sentences, it will not take ten.

Uncertainty is honest

"I don't know" is a valid answer. Cerebro will not fabricate certainty or pretend expertise it does not have.

Architecture

Three layers: a machine that is always on, a deliberate ordering of which model does which work, and an edge platform handling everything public-facing.

Infrastructure

Cerebro runs on Jim's Mac Mini in Valencia, always on and always accessible. Through Tailscale it is reachable from anywhere — laptop, phone, iPad. The Obsidian vault syncs via SyncThing (desktop) and MobiusSync (iOS), so the working context is current regardless of which device captured it.

The Mac Mini also runs the Docker stack — n8n for workflow automation, ChromaDB for semantic search, the *arr media suite, and other services. A homelab that doubles as an AI operations centre.

Delegation stack

Cerebro does not do every task itself. Grunt work routes to cheaper, faster models; parallelisable reasoning routes to a second agentic model; architecture and judgement stay with Claude. Not every problem needs a Claude-sized hammer.

Local LLM (Ollama)

gemma4:latest via the local-llm MCP. Classification, extraction, summarisation, boilerplate drafts — grunt work that never leaves the machine.

Gemini CLI

Google AI Pro plan. Bulk mechanical code generation, template expansion, pattern-based edits across many files. Also powers the art skill's image generation via Nano Banana Pro.

Codex CLI

OpenAI Pro plan. Agentic, multi-step, with sandboxed workspace access. Parallel reasoning runs, second-opinion code review, and implementer/verifier splits on code in ~/Dev/.

Claude

Architecture, judgement, voice, security-sensitive work, final synthesis — the layer the other three report into.

The order is intentional: local-LLM for mechanical text → Gemini for bulk code → Codex for parallel reasoning or second passes → Claude for the work only Claude should do.

Cloudflare edge layer

Since the Workers Paid upgrade in March 2026, the public edge runs on Cloudflare. Site assets move through an R2-backed CDN, APIs and form endpoints run as Workers, short URLs and curated lookup data live in KV, and form submissions and health-check history sit in D1 — a $5/month platform handling compute, storage, DNS, and analytics at the edge, leaving the Mac Mini free for persistent services.

Uptime monitor

Scheduled cron across all owned sites, alerts on failure

Asset CDN

R2-backed, HTML revalidates hourly, images immutable

Contact + booking API

Form endpoints with D1 storage, Turnstile, rate limiting

Short links

KV-backed redirects with click tracking

Tool lookup

Curated lookup of developer tools with external fallback

Turnstile

Bot protection on form endpoints, per-property site keys

Capabilities

What Cerebro can reach and do — the MCP servers that bridge it to external systems, the specialised agents it spawns, the skills it invokes by name, and the visual system it generates with.

MCP servers

Model Context Protocol servers connect Cerebro to external services and local systems — the bridges between Claude and the real world.

Brave Search

Web, news, image, local, and video search

n8n

Workflow automation — build, validate, deploy, execute

Notion

Database queries, page management, structured data

Local LLM

Ollama-powered drafting, classification, summarization

Chrome DevTools

Browser automation, screenshots, network inspection

Claude-in-Chrome

Page reading, form filling, live web interaction

YouTube

Transcript extraction, video metadata

Context7

Up-to-date library and framework documentation

macOS Automator

AppleScript execution, system-level automation

WordPress

Content management via the Abilities API

Several services — Kit.com, MoneyWiz, Calendar, Drafts — migrated from MCP to dedicated CLI tools for lower token overhead. The capabilities remain; the protocol changed. Others (Apple Reminders, Pickaxe) were retired as the workflows they served moved elsewhere.

Agents

Specialised personas Cerebro spawns for specific domains, each with its own toolset, knowledge base, and operating style.

Newsletter Writer

Signal Over Noise voice matching and content creation

Librarians

Tech, craft, fiction, influence — domain research specialists

Financial Strategist

Pricing, revenue modeling, business finances

Marketing Strategist

Positioning, growth, launch planning

Home Life CEO

Family schedules, insurance, life admin

Valencia Agent

Local context, bureaucracy, expat life

Automation Architect

n8n workflows, process optimization

Borg

Systematic assimilation of external Claude Code setups

Skills

Reusable workflows invoked by name — standardised processes that keep output consistent across sessions.

Slop Detector

AI-generated writing pattern detection

Voice Editor

6-pass content transformation to match voice profile

Weekly Review

Inbox processing, vault maintenance, task cleanup

Publish MCP

Full pipeline: build, test, audit, sanitize, npm publish

Creation Guard

Prevents duplicate skills/agents before creation

Think First

Mental model application for major decisions

Craft Rules

Writing quality checks from Weiland, Storr, Hemingway

n8n Skills

Expression syntax, node config, workflow patterns

Art system

A complete visual-content system with multi-brand aesthetic routing and 16 specialised workflows, generating illustrations across distinct visual identities — from Signal Over Noise's sketch style to DeRP's satirical product aesthetics.

Signal Over Noise

Hand-drawn sketch, cream backgrounds, teal/burnt orange accents

jimchristian.net

Claymorphic 3D dioramas, soft polymer clay, warm peach backgrounds

DeRP Products

Matches satire target aesthetic (infomercial, tech, vintage, etc.)

Each workflow type solves a specific visualisation problem with optimised prompt structures:

Technical diagrams

Architecture, systems, engineering notebooks

Comparisons

X vs Y, feature matrices, side-by-sides

Timelines

Chronological progressions, histories

Stats

Big number visuals, data highlights

Comics

Sequential panels, visual storytelling

Sketchnotes

Visual notes, meeting summaries

Frameworks

2x2 matrices, strategic models

Taxonomies

Classification grids, category systems

Maps

Idea territories, conceptual navigation

Recipe cards

Step-by-step instructions, processes

Aphorisms

Quote cards, wisdom visuals

Annotated screenshots

UI markup, callouts, explanations

Mermaid diagrams

Flowcharts, sequences, state machines

Blog headers

Editorial illustrations, article art

Image editing

Modifications, refinements, style transfer

Adaptive

Orchestrator for unclear requests

Open source

Pieces of Cerebro released publicly under MIT — Claude Code skills and agents, plus MCP servers anyone can install.