Joyput
A platform for building persistent AI systems.
View live project →/01 Context
AI applications keep rebuilding context for every model and vendor. Joyput exists to fix that — a platform where agents and operators share one durable substrate: memory, runbooks, tickets, panels, and hosted MCP inside an isolated Space.
I founded Joyput to build persistent AI systems at the intersection of software engineering, product design, and artificial intelligence. The goal: let AI clients operate from the same persistent foundation instead of reinventing context plumbing in every tool.
/02 The Build
The Engine
I built the core runtime myself — gateway routing, MCP hub, semantic memory pipeline, and a host provisioner that spins up isolated Spaces on demand. Each Space is a tenant boundary: its own MCP endpoint, memory store, runbooks, and panel surfaces.
The hub delegates to Space gateways so external clients like Cursor and Claude can retrieve memory, run runbooks, and operate inside the correct scope — without the dashboard or the agent guessing which context applies.
The Dashboard
The full-stack dashboard is a Next.js application: authentication with pilot gating, a Connect AI wizard for hub tokens, Control Central for admin operations, composable panels, and Space management — create, provision, invite collaborators.
I owned every layer: API routes, database schema, email flows, onboarding, and the operator UX. From waitlist approval to first MCP connection, the product path is designed end-to-end.
Infrastructure
Production runs on a two-droplet layout — dashboard and core — with Caddy reverse proxy, Cloudflare TLS, Postgres, systemd services, and managed Space provisioning over a private VPC.
I wrote the deploy runbooks, smoke tests, and reconciliation scripts. Pilot go-live included OAuth, Resend email, hub membership verification, and external MCP client connection — all shipped and verified on prod.
/03 Under the Hood
Joyput combines a custom runtime (joyput-core) with a Next.js dashboard (joyput-clients). Spaces get per-tenant MCP URLs, semantic memory ingestion, runbook execution, ticket tracking, and panel widgets — all reachable through hosted MCP.
The MCP hub mints tokens, verifies membership against the dashboard, and proxies tool calls to the active Space gateway. Host provisioning creates isolated stacks on the core droplet; the dashboard orchestrates lifecycle from a single wizard.
/04 Outcomes
Joyput is live at joyput.com with a production pilot: user signup, waitlist gating, Space provisioning, hub MCP connection, memory retrieval, panels, and collaborator invites — all verified end-to-end.
This is the project where design, engineering, and AI infrastructure converge. I built the engine, the dashboard, the provisioning layer, and the deploy pipeline — everything.
"Stop rebuilding context plumbing in every tool. Joyput gives agents and operators the same durable substrate."
James Carvallyo / Founder