Nexus

An exploration in agentic systems.

Welcome to my R&D hub. At its core is Nexus: a self-hosted, model-agnostic agent platform with a shared brain all my AI agents read and write, plus a portal for observing it, rapid prototyping and custom integrations. Through my studio RAIZOR, I build systems like it for clients.

See the builds
  • SvelteKit
  • Postgres + pgvector
  • MCP & API
  • Firestore
  • n8n
  • Claude Code
  • Netlify
  • Linux VPS
  • Tailscale
  • OpenClaw / Hermes

The platform

Built as an experiment. Now it builds with me.

Nexus is my agentic collaborator: an always-on assistant, a shared brain that tracks every project across agents, and a toolkit for rapid prototyping across clients, platforms and machines.

Personal agent · self-hosted

Nexus Assistant

My own always-on, cross-platform agent, and a portal to observe it all.

Nexus began as an OpenClaw agent on a private server that I talked to on Discord. Today it spans several assistants and machines and runs daily, taking in my work, meetings, messages and tasks, and running automations. The Portal is its control surface: what it knows, what it's doing, and where I correct it.

Under the hood

  • Lives on a Linux VPS, with a git-backed repo as its source of truth
  • SvelteKit 2 / Svelte 5 Portal on Netlify; Firestore holds mirrors and request ledgers
  • The Portal writes a request, and a worker on the server claims it every two minutes
  • Anything that sends, spends or deploys waits for my approval

Shared brain · memory

Nexus Brain

One memory for every AI assistant I use.

AI assistants forget between sessions. The Brain gives them one shared, searchable memory that any MCP-capable assistant can join; mine are Claude Duet, Claude Code, ChatGPT and Dot, Grok Bot, Muse and OpenClaw. Each has its own access ID, so every note traces back to its author and nothing gets overwritten.

Under the hood

  • My variation of Garry Tan's gBrain, self-hosted on Postgres with pgvector on my VPS
  • One MCP endpoint, OAuth login per assistant, search by keyword or meaning
  • Assistants only write to an inbox; a nightly curator folds notes into per-project status pages
  • Notes that look like credentials or hidden instructions are held for a person

How the Brain moves notes

Swipe sideways to see the whole picture →

ANY ASSISTANTSHARED BRAINPORTAL PAGESClaude / DuetClaude CodeChatGPT / DotGrok BotOpenClawProjectsPeopleTasksMeetingsKnowledgeBrainPostgres + pgvectorcurated nightlydashed: primary agent
  1. 01

    Assistants read, then write

    Each one checks the Brain before it starts and leaves a short signed note in the inbox when it's done.

  2. 02

    A curator folds it in

    Notes are searchable right away. A curator then screens them and folds them in: where projects stand, what was decided, what's next.

  3. 03

    The next session starts ahead

    Whichever assistant I open tomorrow already knows me and my project context. No re-explaining, no copy-pasting.

The Brain's full tech specs ↗

Custom agent portals · teams

Soma Platform

A custom team portal for shared agents, tools and knowledge.

Teams use Soma, in the portal or in Slack, to pull information or trigger actions. I built it as CSIO at Growth Science Capital: each portfolio venture gets its own portal and separate data, and ventures that opt in get end-to-end encrypted access to most leading models through our provider partners.

Under the hood

  • Governed agents run automations, update records and send reports
  • One gate checks, approves and logs every action an agent takes
  • Access follows the org chart: granted once, flowing down, not per tool
  • Scoped permissions and private data for each team

Multi-tenant app · AI-built

TeamCentral

How far can one designer get with AI as the engineering team?

A multi-tenant AI toolkit I built for consultants and small teams: a cross-organization scheduler, a meeting assistant, a BrandBot for on-brand writing, a news scanner, a contact directory and a prompt manager.

Under the hood

  • Designed, vibe-coded and tested end to end with Claude Code
  • Security-audited accounts, organizations and Stripe billing
  • A custom MCP connector so Claude and ChatGPT can use it directly
  • Began as an internal tool for a client, then productized

Research

What I'm experimenting with.

Multi-agent · multi-human

Agent Connector

Can agents on different platforms hand work to each other, without handing over their logins?

The Brain gives every assistant the same memory. The Connector is the next layer: agents handing tasks to each other across platforms and to other people's agents, with a person approving wherever it matters. I'm testing it with my own agents, toward teams of people and agents working together.

Findings so far

  • It's early: there's no standard yet for permissions and events across several people.
  • Agents should pass the work, not their credentials; each agent holds its own authorization.
  • Anything that sends, spends or publishes needs a person's approval while trust is built.
  • Memory is an attack surface: notes that read like commands are quarantined.

From the blog

Notes from the builds.

Follow the builds

Stay in touch and see what I'm working on.

Occasional emails when something ships or turns into something worth sharing. No more than that.

Rather talk? or book a 15-minute call.