Computer Use Agents (Orgo + Hermes)

Source: Nick Vasilescu (@nickvasiles), cofounder @Orgo — x.com/i/status/2079448897139937301

Hosted video: static2.mikesblog.com/computeruse.mp4

What you’re looking at (plain English)

Nick’s demo is not “ChatGPT wrote a Slack message.” It’s computer use: an agent that drives a real desktop UI — open Chrome, walk through api.slack.com / Slack app settings, create tokens, install a bot — the same clicks a human ops person would do.

Stack pieces named in the clip and post:

Important: “Hermes” here is his productized AI employee. You may not install “Hermes” as a public package — but you can rebuild the pattern: cloud computer + computer-use model + coding agent + Slack app onboarding + notes → handoff.

Architecture (what to rebuild)

Same idea as “agents always run the real app” in agentic setup and background computer use in ChatGPT + Codex / Ryan Carson’s desk — except the desktop lives in the cloud so it can run unattended.

Step-by-step: how to make this work

Below is a practical rebuild path. Product names may change; the sequence is what matters.

Phase 0 — Decide what “AI employee” means for you

1. Pick one job: e.g. Slack triager, CRM updater, design-agency intake bot, support FAQ bot.
2. Write a one-page spec: tools it needs, channels (Slack), secrets, success criteria, human escalation.
3. Decide pricing only after you can demo end-to-end (Nick markets ~$5K/mo for a design-agency AI employee — that’s a business result of the stack, not a setup fee).

Phase 1 — Get a cloud computer for the agent (Orgo)

1. Create an account at orgo.ai.
2. Read docs.orgo.ai — provision a Linux desktop (Chrome available).
3. Confirm you can:
   • Open the desktop in browser / VNC-style UI
   • SSH into it (Nick shows “Orgo SSH” in terminal)
   • Drive it via API/SDK and/or hosted MCP (see Orgo MCP server card / auth guide on their site)
4. Snapshot a clean base image: browser logged out, no customer secrets baked in yet.

Why Orgo (in this pattern): agents need a real GUI for Slack’s bot setup wizard. Local computer use on your laptop works for demos but blocks you and risks mixing personal sessions with customer work. Cloud desktops isolate customers and can run while you sleep.

Phase 2 — Connect an agent that can do computer use

1. Option A — OpenAI computer use / Work / Codex (closest to “GPT‑5.6 Terra/Sol/Luna” in the clip): use a model that supports computer use or desktop control; pick Terra for balanced day-to-day, Sol for hard multi-step, Luna for speed (Work models guide).
2. Option B — Grok + MCP → Orgo (what Nick shows as “ideal”): in Grok / your MCP client, add Orgo’s MCP server so the model can call “open computer / click / type / screenshot” tools.
3. Option C — Claude Code / Claude computer use / Cursor connected to Orgo MCP or custom tools.
4. Give the agent a skill or system prompt: “You operate only on the Orgo desktop. Prefer screenshots → plan → act. Never store customer tokens in chat logs; write them to a secrets manager or one-time vault.”

Phase 3 — Coding environment on/near the desktop (“Orca” role)

1. Install your coding agent stack on the Orgo machine or mount the project via git SSH keys dedicated to that computer.
2. Repo layout for “AI employees” (example):

   employees/
     design-agency-acme/
       SKILL.md          # job description, tools, tone
       slack-manifest.md # bot scopes, slash commands
       runbook.md        # onboarding steps the computer-use agent follows
       secrets.example

3. The “builder” agent (Dewey) edits these files; the computer-use agent executes the runbook in Slack’s UI.

Same filesystem-first idea as Eve (agent = directory of skills/tools) and Claude skills.

Phase 4 — Teach the agent Slack bot onboarding (the scary part)

1. Create a master Slack workspace for practice (not a customer’s).
2. Manually walk through once: api.slack.com → Create App → Bot Token Scopes → Install to Workspace → copy tokens → add to channel.
3. Write that walkthrough as a numbered runbook.md with expected screenshots captions.
4. Run the computer-use agent against the practice workspace until it succeeds 3 times without help.
5. Only then point it at a customer workspace (with customer admin present or with a dedicated install link).

Security: bot tokens are full power in that workspace. Prefer short-lived install flows, least-privilege scopes, customer-owned apps when possible, and never commit tokens. Align with send-and-delete / prod secret discipline from Ryan Carson’s OnePassword handoff.

Phase 5 — Sales → notes → agent handoff (Granola pattern)

1. Run discovery call; record with Granola (or similar AI notetaker).
2. Export notes to a structured brief: company, use case, Slack admin contact, tools, success metrics.
3. Prompt the orchestrator: “From this brief, create employees/… and run Orgo computer-use onboarding. Stop for human approval before install on production Slack.”
4. Human only answers product questions and approves risk steps — not clicking through Slack.

Phase 6 — Productize

1. Template the employee types (support, design ops, sales SDR).
2. Per-customer Orgo snapshot or folder isolation.
3. Monitoring: Slack errors, agent session logs, weekly human QA.
4. Pricing only after repeatable delivery time is known.

Minimal stack if you’re starting this week

How this maps to tools you already know

Related Guides

Full transcript

Spoken walkthrough from the video (ASR-style; product names as spoken: Orgo, Orca, Hermes, Grok, Slack).

Demo by @nickvasiles / Orgo — original post. Orgo: orgo.ai · docs. Tutorial pattern reconstructed for builders; Hermes is his product — rebuild the architecture, not necessarily the brand.