Monkey Agentic OS
An agentic AI platform for my web-app teams' workflows.
New Routing →
session 1
session 2
session 3
Terminal ↗Documents ↗Agents work tree ↗
orchestrator
leadopus · high
working412k · $8.10
ui-frontend2
ui-frontend-lead
leadopus · high
working96k · $1.90
frontend-builder
membersonnet · med
working128k · $2.40
+ create agent
seo2
seo-lead
leadopus · high
working74k · $1.45
seo-technical
membercodex · low
working41k · flat
+ create agent
backend-data2
backend-architect
leadopus · high
error141k · $2.90
qa-engineer
memberhaiku · low
idle · waiting9k · $0.05
+ create agent
research2
deep-researcher
leadsonnet · med
working188k · $1.10
research-users
membersonnet · low
finished31k · $0.20
+ create agent
orchestrator vicente-portfolio main This Mac
Plan, build, / for skills, @ for context
Sonnet · Medium
- 46agents, orchestrator included
- 13clusters
- 39skills
- 8workflow commands
- 4UI themes, AA each
What it is
I design the product. Agents help me build it.
Monkey Agentic OS runs my web design and development, SEO and research work.
I act as product and UX lead: I write the specs, set the rules and review every result. The agents do the scoped work in between.
It is built for my own work, and for the teams I assemble on each project.
Research
How I decided what to build.
From problem to specs, in seven steps. Pick one.
- 15research reports
- 53decisions logged
- 25feature specs
Why spec-driven: spec → UI → design system → backend
- Design goalMore output per session in AI dev environments.
- Design goalFewer tokens spent on later changes.
- Design goalLess time fixing and bolting on features.
Build
Design and code, in parallel.
AI coding made designing and testing this cheap to iterate for me.
Figma
Shared tokens
Code repo
Architecture
One orchestrator. Thirteen specialist clusters.
Pick a cluster on the board. The thread shows its agents.
Threadsession 1
Plan, build, / for skills, @ for context
Routing
Every task goes through one gate.
Pick a task to see its cluster, owner and model.
orchestrator
reads INDEX.md only
How it thinks
Four rules every agent follows.
Read only what the task needs
Context is RAM. Agents open a routing index first, then the few files that matter.
- clusters/INDEX.md
- ui-frontend/INDEX.md
- seo/…
- research/…
- branding/…
Right model, right effort
Effort is a dial too: low for mechanical work, high for hard calls.
- fable
- opus
- sonnet
- codex
- haiku
One owner per file, then review
Parallel agents never touch the same file. Each build passes review and an accessibility audit.
Memory between sessions
Every workstream ends with a state file under 50 lines. The next session starts there.
00-ESTADO-Y-RETOMAR.md
✓ Done · last session first
… In progress
→ Next literal action
! Decisions and gotchas
Design system & UI
Every value is a token. Every theme passes AA.
- Color, spacing, radius, motion and z-index all tokenized.
- Indigo, black, light and custom themes share one token set.
- Separate UI package on shadcn, with Phosphor icons.
- Figma library and code kept in sync.
- Built with my Design System OS: 64 component specs, 17 layouts.
Same system. Three themes.
Pick one. The whole page changes with it.
Screens
Four views of the console.
Live tree of who is working, on which model, at what cost. From the Figma file, rebuilt in code.
New screen, same tokens: spend per provider with search, filter and sort.
Spec-driven boardSpec → UI → Design system → Backend
Demo data
Brainstorm · spec-driven
Brainstorm
fable · high
No UI until the spec is approved
Plan, build, / for skills, @ for context
Specs come out of a brainstorm with the agent. Nothing ships unapproved.
Monkey Agentic OS
Sitemap
Demo data
Canvas. Drag or use arrow keys to pan, plus and minus keys to zoom.
100%
Describe a page or section…
Sitemap canvas with sections per page, generated from the product spec.
CLI Planned · in development
One command: vhs.
Coming: the same agents and context, from any terminal.
ProblemsOutputDebug ConsoleTerminalPorts
vhs · v0 concept
$ vhs
Monkey Agentic OS v0 · concept
opus · high · orchestrator
@orchestrator · ~/vicente-portfolio
❯ audit the pricing page: SEO, a11y, conversion
● Split into three owners, in parallel:
Read clusters/INDEX.md
→ seo-lead
→ a11y-auditor
→ cro-auditor
Ran 3 subagents
❯ Plan, build, / for skills, @ for context
▸▸ auto mode on (shift+tab to cycle)
SubscriptionIf the claude CLI is installed, it runs on my Claude subscription.
APIsIf not, it falls back to provider API keys set in settings.
MCP
Plugged into the tools I work in.
Connected through MCP, CLIs and APIs. More can be added.
- Spline
- Screaming Frog
- Apify
Project facts
Role, stack and status.
- Role
- Product & UX designer and builder. Built for me and the teams I assemble per project.
- Stack
- Claude Code, Codex, Next.js, pnpm monorepo, shadcn, Phosphor, Figma.
- Status
- Agent system in daily use. Console UI in build, not released. CLI (
vhs) planned.
Want to see it run?
I can walk you through a live session.