No-Shell Agent Operating Pack

An open-source operating pack for people who already use AI agents but keep getting "empty-shell automation": files, buttons, or dashboards that look finished but do not actually run a real workflow.

This is not a prompt-pack that promises magic. It is a plain-language operating system for delegating work to ChatGPT, Claude, Codex, Gemini, Hermes, and similar agents with a real objective, inputs, tool choice, validation, recovery, and human boundary.

No-Shell Agent Architect MCP

This repository now also ships a remote MCP/API product: No-Shell Agent Architect MCP.

It takes a plain-language automation goal and returns:

the recommended MCP/plugin/skill/tool stack;
a copy-paste no-shell agent command;
dry-run steps;
QA and audit checks;
recovery guidance;
human boundaries for secrets, accounts, payments, publishing, and irreversible actions.

Local API preview:

npm install
npm run dev

Then open:

http://localhost:3000

Remote MCP endpoint after deployment:

https://no-shell-agent-architect-mcp.vercel.app/api/mcp

Live demo:

https://no-shell-agent-architect-mcp.vercel.app

Launch kit:

https://no-shell-agent-architect-mcp.vercel.app/launch

Official MCP Registry:

https://registry.modelcontextprotocol.io/v0/servers?search=no-shell-agent-architect-mcp

Client config:

{
  "mcpServers": {
    "no-shell-agent-architect": {
      "url": "https://no-shell-agent-architect-mcp.vercel.app/api/mcp"
    }
  }
}

Who this is for

Non-developers, solo operators, freelancers, and small teams who want practical AI automation.
People who know AI can code or automate, but do not know which plugin, skill, command, or workflow to start with.
People who need "make this business workflow actually work", not only "generate some code".

What's included

Free diagnostic: 10 reasons AI automations become empty shells.
Starter pack: Email and document automation workflow.
Before/After demo: a shallow command vs a no-shell operating command.
Customer intake sheet.
Tool/router sheet for choosing agents, plugins, skills, MCP, browser, documents, spreadsheets, or manual review.
Copy-paste command cards written in natural language.
Result scorecard for catching fake completion.
Recovery playbook: fix the upstream cause instead of stopping at one more gate.
Validation tracker for a 20-person usefulness test.
Local landing mockup, PDF exports, and ZIP packages.

Quick start

Read free/ai_automation_failure_diagnostic_card.md.
Open demo/before_after_email_document_demo.md.
Use delivery/01_customer_intake.md to describe one real workflow.
Pick the tool path in delivery/02_tool_router.md.
Run the command card in delivery/03_command_cards.md.
Check the result with delivery/04_result_scorecard.md.
If it fails, use delivery/05_recovery_playbook.md.

Rendered artifacts

Local landing mockup: landing/index.html
Free PDF: pdf/free-diagnostic-card.pdf
Demo PDF: pdf/before-after-demo.pdf
Customer ZIP: dist/no-shell-agent-operating-pack-starter-v1.zip
Workbench ZIP: dist/no-shell-agent-operating-pack-workbench-v1.zip
Validation tracker: outreach/validation_tracker.csv

Validation

Run the package selfcheck:

npm run selfcheck
npm run test
npm run build

Expected result:

PASS package selfcheck
PASS mcp product selfcheck

The selfcheck verifies required files, rendered PDF/PNG artifacts, ZIP contents, landing copy, README/manifest, validation-tracker fields, MCP metadata, tool coverage, and secret-pattern safety.

Marketing

Launch copy, community drafts, Product Hunt copy, Show HN copy, Reddit/LinkedIn/X drafts, and directory submission text live in docs/LAUNCH_KIT.md.

Current status

This repository is public and open source. It is early-stage and still needs real external validation. The local package is complete enough to show to users, but market demand is not proven until the 20-person validation tracker has real responses.

No payment connection, storefront publish, account login, secret, or private customer data is included in this repository.

License

MIT. See LICENSE.