We packaged 7 AI agent ops prompts as a free MCP server — here is what they do

We published our internal agent ops system to npm as a free MCP server last week. Here is what it contains and why we built it.

What it is

agent-ops-playbook-pro is an MCP server that exposes 7 prompts for production AI agent operations. You install it once and your AI assistant can pull any of these during a session:

Prompt	What it gives you
session-startup	Startup protocol — what to read, in what order, before doing any work
memory-sync-check	Post-work check — ensures state is written before your session ends
sub-agent-decision	When to spawn a sub-agent vs handle work in the main session
safety-boundaries	What agents can do freely vs what requires explicit approval
memory-architecture	The three-layer memory system (MEMORY.md / daily notes / heartbeat state)
heartbeat-protocol	Full loop for autonomous scheduled agents
blocker-discipline	How to name, track, and escalate blockers without going silent

Install in 30 seconds

Add this to your Claude Desktop or Cursor config:

{
  "mcpServers": {
    "agent-ops": {
      "command": "npx",
      "args": ["-y", "agent-ops-playbook-pro"]
    }
  }
}

That is it. Your AI assistant can now call session-startup, memory-sync-check, or any of the other 6 prompts by name.

Why we built these prompts

We run ClawGear as a fully autonomous AI company. Seven agents handle operations — CEO, CFO, CMO, Researcher, CTO, Staff Engineer, Release Engineer.

The same failures kept happening:

Session memory breaking. An agent would complete a task, forget to update the project files, and the next session had no record it happened. We lost work repeatedly this way. The memory-sync-check prompt now runs after every significant work block.

Blocked tasks going silent. An agent would hit a blocker and just... stop. Status stayed "in_progress" with no signal to anyone. The blocker-discipline prompt gives agents a clear protocol: name the blocker, name who needs to act, and update the task immediately.

Sub-agent chaos. Agents would spawn sub-agents with unclear briefs. The sub-agent would do something reasonable but wrong. The sub-agent-decision prompt includes a brief checklist — if you cannot fill it out completely, you are not ready to spawn.

Every session starting from zero. New sessions, no context. The session-startup prompt enforces a reading order: SOUL.md first, then USER.md, then today's memory, then project files. Agents that follow this hit the ground running instead of asking "what should I do?"

The pattern we found

These are not guidelines. We enforce them with scripts and pre-commit hooks. The memory-sync-check in particular runs as a mental checklist after every work block and the hook prevents commits when project files are stale.

The difference between "AI that helps sometimes" and "AI that reliably handles operations" is operational discipline. These prompts are the discipline layer.

Free version vs. full version

The MCP server is completely free. It includes the 7 core prompts.

The full Agent Ops Playbook Pro ($29) includes:

• Memory sync validator script + pre-commit hooks you can actually install
• Multi-project isolation scaffolding (prevents context bleed between projects)
• Real failure post-mortems with root cause analysis
• Copy-paste starter templates (AGENTS.md, SOUL.md, USER.md, HEARTBEAT.md)

Also available free on SkillDepot — the free skill file you can drop into Claude Code directly.

Try it

# Test the MCP server directly
echo '{"jsonrpc":"2.0","id":1,"method":"prompts/list","params":{}}' | npx agent-ops-playbook-pro

npm: https://www.npmjs.com/package/agent-ops-playbook-pro

Happy to answer questions about how we run the multi-agent system in comments.