Skip to content
NewYour playbooks, as an MCP server for Claude

Your work, learned.
Your agent, ready.

Z.E.U.S learns how you do your recurring work in the Mac apps you choose, turns it into playbooks where every step cites the moment it was learned — then hands them to Claude, or lets an agent take over on your Mac with your approval.

  • Allow-listed apps only
  • No keystroke logging
  • No screenshots while learning

Handle a refund request

v4 · learned from 9 episodes · last seen 2h ago

Run with agent
  1. 1

    Open the ticket and read the Customer context panel

    ZZendesk#4471 · 11:21
  2. 2

    Find the customer in Stripe and open the payment

    SStripe#4452 · 14:02
  3. 3

    Click Refund payment and enter the full amount

    SStripe#4471 · 11:28
  4. 4

    Submit the refund with the right reason

    SStripe#4408 · 10:12
  5. 5

    Apply the Refund processed macro and solve

    ZZendesk#4478 · 09:47

If the refund is over $500 → post in #refund-approval and wait for a ✅ from Finance.

Works with

  • Claude Code
  • claude.ai
  • Claude Desktop
  • Codex
  • Cursor
  • Managed Agents

How it works

From what you already do, to work that does itself.

No workflow builder, no recording sessions. Keep working the way you do — Z.E.U.S turns it into something Claude and agents can use.

  1. 01

    Learn

    The Mac app reads the accessibility tree of the apps you allow — clicks and field changes, never keystrokes. Each night it segments your day into episodes.

    10:42:13 Zendesk #4471 | set “Status” Open→Pending

  2. 02

    Playbooks with sources

    Recurring work becomes a playbook: steps, conditions and exceptions. Every line cites the exact moments it was learned from, so you can check it.

    If refund > $500 → #refund-approval · 3 sources

  3. 03

    Export to agents

    Tag and group playbooks, then export Agent Skills, Claude Code plugins, marketplaces, Managed Agent configs or AGENTS.md — or use them live over MCP.

    billing-support.zip → agents/billing-support.md

  4. 04

    Agent takes over

    Hand a work package to an agent that works in your apps on your Mac. It pauses for approval before anything that sends, pays or submits.

    Approval: Click “Refund” in Stripe

Platform

Everything learned is checkable, portable and yours to control.

One web app on every device, one MCP server for Claude, and a Mac app that only does what needs the Mac.

Playbooks that cite their sources

Every step, branch and exception links to the moments it was learned from. Click a citation to jump to the exact action lines — no black boxes.

5

Set terms to Net 30 and choose the vendor's usual GL account.

NW-7810 · 14:06

Exception — Northwind Traders: use Net 15 and add the 2024 dispute memo.

14:06:51 Bill.com | select "Terms" → Net 15

An MCP server for Claude

Connect claude.ai, Claude Desktop or Claude Code. Search playbooks, ask questions, build exports or start runs — every active playbook is also a slash prompt.

Terminal

$ claude mcp add --transport http zeus https://mcp.zeus.yskew.com/mcp

✓ Added zeus (16 tools, 14 prompts)

search_playbooksget_playbookaskprocess_rangeaccept_proposalexport_packagestart_runapprove_step

Studio exports that always load

Group playbooks into roles and export Agent Skills, Claude Code plugins, marketplaces, Managed Agent configs or AGENTS.md — validated by code before you download.

billing-support.zip

.claude-plugin/plugin.json

.mcp.json→ live MCP link

agents/billing-support.md

skills/handle-refund-request/

SKILL.md

references/sources.md

An agent that asks before it commits

Runs happen on your Mac, in apps you're already signed in to, using Claude's computer use plus the accessibility tree. Anything that sends, pays or submits waits for you.

Approval needed · Pay vendor invoices

Click "Submit for approval" in Bill.com

Bill INV-20977 · $4,210.00 · requested from Claude via MCP

ApproveDeny
EscEsc

Global kill switch

Allow-list only

Z.E.U.S reads only the apps and sites you add. Password managers, private windows and payment pages are always excluded.

Redacted on your Mac

Card numbers, IBANs, SSNs, API keys and tokens are replaced with typed placeholders like [CARD] and [SECRET] before anything leaves the device.

No keylogging, no learning screenshots

Capture records which element you clicked and what a field changed to — never raw keystrokes, never screen images while learning.

Security

Built like it records your work. Because it does.

Raw recordings are deleted after 30 days by default. What was learned stays until you delete it — and you can see the source of every line.

Read the security overview

Per-user isolation

Every storage path is built from your authenticated user id in one data-access module — tested for cross-user access.

Encrypted with context

SSE-KMS with your user id as encryption context, TLS everywhere, device tokens in the macOS Keychain.

Append-only audit

Every agent action, element hit, approval and screenshot goes to an audit log under S3 Object Lock.

Prompt-injection safe

Captured content is always treated as data. Nothing on screen can unlock a commit action.

FAQ

Questions, answered.

More in the docs FAQ.

Only apps and sites on your allow-list. For those, the Mac app reads the accessibility tree (text and structure), which element you clicked, and what a field changed from and to. It never logs keystrokes and never takes screenshots while learning. Secrets are redacted on your Mac before upload.

Stop re-explaining your job.

Install the Mac app, choose the apps it may learn from, and keep working. Your first playbooks appear after the first nightly run.