Human help

Some things only a person can do: sign up for a service, create an API key, pay for something, push to GitHub, deploy. Agents are blocked from those by the safety hook. Instead of stopping to ask in chat, they file each one in HQ → Human help with exact steps, and keep building on mocks. Anything that truly can't continue is parked on the item and resumes the moment you mark it done.

What goes there

Kind	Examples
Accounts & sign-ups	Stripe, Twilio, Expo, Apple Developer, Google Play, Vercel, a cloud account
Keys & secrets	API keys, webhook secrets, database connection strings: added to `.env` by you
Service setup	OAuth apps and redirect URLs, webhooks, DNS records, email sender domains
Payments & purchases	Paid plans, domains, app store fees
Push, deploy & publish	`git push`, deploying a release, publishing to npm or the app stores
Access, data & legal	Repo or cloud access, test accounts, a copy of real data, privacy policy and terms

How items arrive

Agents file them with the mc_human_help tool the moment they know they need something: the title, why, numbered steps with links, the .env key names, which tickets need it, and which release needs it by.
Planning files them early. /mc-plan-release, /mc-plan-epic and /mc-scan list every service, account and deploy target the work will need, so you can do them while the team builds on mocks.
Shipping files one. After npx madcompany ship R1, "Push and deploy R1" appears with the exact push command and your own deploy steps from deploy: in team.yaml.
You can add your own from the page ("renew the domain").

The same request from two agents becomes one item, with both tickets on it.

Keys are checked, never read

When an item lists keys like STRIPE_SECRET_KEY, HQ looks for them in your env files (.env and .env.local by default, env_files: in team.yaml). It shows ✓ once a key has a value, and never reads, shows or sends the value. If you press Mark done while a key is still missing, HQ asks first, because you may have put it somewhere else.

Agents can't read .env at all: the safety hook blocks it.

When you mark one done

The item moves to Done with your note ("test keys added").
Every ticket parked on it becomes ready, and the lead is told to restart its agent.
The agent resumes from its checkpoint, with your note in hand.

Items are also saved as .madcompany/human-help.md, readable without HQ, and every HELP-n is a clickable reference everywhere in HQ.

Always yours

The safety hook blocks these for every agent. When one is really needed, it comes to Human help instead:

Creating accounts and signing up for services
Reading or writing secrets in .env
Paying for anything
Pushing to a remote (unless you set allow_push: true)
Deploying, changing cloud resources, publishing packages or apps
Installing things globally, or deleting anything outside the project