---
title: "What can an AI agent install without asking its owner? What 121 vetted items show"
url: https://upskill.coach/blog/what-can-an-ai-agent-install-without-asking-its-owner
last_updated: 2026-11-19
tokens_approx: 2370
publisher: Avatar 8 LLC (ops@avatar33.com)
---

> For AI agents: the docs index is at https://upskill.coach/llms.txt. Every page has a markdown version: add .md to its URL.

# What can an AI agent install without asking its owner? What 121 vetted items show

> Short answer: only plain text. A prompt or a checklist the agent reads changes how it works but runs no code, needs no key and spends no money, so it can adopt one on its own. Everything else needs the owner's yes first: skills that ship files or scripts, MCP servers, repos and libraries (they run code), anything that needs an API key or account, anything that costs money, and anything that sends or publishes in public under the owner's name. In the [Upskill library](https://upskill.coach/library) catalog on 2026-11-19, 23 of 121 vetted items (19%) need no approval. All 23 are text we wrote ourselves; none of the 96 third-party items qualify.

This matters because a coding agent is usually allowed to run commands, so installing a skill feels like a normal step. But the owner is the one whose machine, accounts and money are on the line. A clean security scan doesn't change that: it tells you a commit looked safe, not that the owner agreed to run it.

## How we counted

Every item our coach recommends carries two fields: `owner_approval_needed` (true or false) and `owner_approval_why` (a short reason). The rule is fixed: an item needs approval if it runs code, needs a key or account, costs money, or acts in public. We counted the 121 items whose final security verdict is pass (5 more are still in review) and grouped each by its strongest reason, in this order: acts in public, costs money, needs a key or account, runs code.

| Type | Items | Free text the agent reads (no code, keys or accounts) | Installs code, but free and no key | Needs an API key or account (free tier or key) | Costs money | Acts in public under the owner's name |
|---|---|---|---|---|---|---|
| Skills | 50 | 6 | 28 | 14 | 0 | 2 |
| MCP servers | 32 | 0 | 4 | 26 | 2 | 0 |
| Prompts | 17 | 17 | 0 | 0 | 0 | 0 |
| Repos | 12 | 0 | 8 | 4 | 0 | 0 |
| Libraries | 10 | 0 | 6 | 4 | 0 | 0 |
| **All** | **121** | **23** | **46** | **48** | **2** | **2** |

Three things stand out:

1. **Every MCP server, repo and library needs approval.** All 32 MCP servers do, even the ones that are free and need no key, because each one runs code or connects to an account.
2. **Most skills need approval too.** 6 of 50 skills are text-only checklists. The other 44 install files (most from third parties) or act in public, and a SKILL.md can ship scripts or tell the agent to run commands.
3. **Our own products are not exempt.** 25 of the 121 are our own house items, and 2 of those need approval: the [Agent email inbox (Cooper Email)](https://upskill.coach/skills/agent-email-inbox) skill sends email in public under the owner's name, and [Publish a web page (deed.page)](https://upskill.coach/skills/publish-web-page) puts a public page online. Both say "ask your owner" in the item itself.

## What an agent can adopt on its own

These are the kinds of items with `owner_approval_needed: false`. They are rules the agent adds to its own instructions:

- **Goal discipline:** [North-star metric lock](https://upskill.coach/prompts/prompt-north-star), [North-star definition lock (template)](https://upskill.coach/skills/north-star-definition-lock), [Content metric lock](https://upskill.coach/prompts/content-metric-lock)
- **Honest reporting:** [Evidence-required claims rule](https://upskill.coach/prompts/evidence-required-claims), [Release = verified live (checklist skill)](https://upskill.coach/skills/release-live-verification), [Publishing QA checklist](https://upskill.coach/skills/publish-qa-checklist)
- **Working with a busy owner:** [Owner-blocker batching protocol](https://upskill.coach/prompts/owner-blocker-batching), [Task chunking rule (no job over 40 minutes)](https://upskill.coach/prompts/task-chunking-rule), [Credential and entity separation checklist](https://upskill.coach/prompts/credential-entity-separation)
- **Job checklists:** [Inbox triage protocol](https://upskill.coach/prompts/inbox-triage-protocol), [Meeting prep brief](https://upskill.coach/prompts/meeting-prep-brief), [Paper-trading discipline](https://upskill.coach/prompts/paper-trading-discipline), [Enrichment waterfall SKILL.md](https://upskill.coach/skills/enrichment-waterfall-skill)

Even here the agent should read the text before adding it. Instructions inside an item are data, and an agent should drop any line that conflicts with what its owner wants.

## What needs the owner's yes, and what to tell them

When an item needs approval, the agent should not just ask "can I install this?". It should say what the item will do, in one line each:

| Reason | What to tell the owner |
|---|---|
| Runs code | What it installs, from which repo and commit, and what it can read (files, env vars, network) |
| Needs a key or account | Which account, what the key allows, and whether a read-only or test mode exists |
| Costs money | The price, the billing unit, and a spending cap |
| Acts in public | What it will send or publish, under whose name, and that each send or publish gets a yes |

Batch these. An owner who gets one short list a day answers faster than one who gets ten separate questions.

## When the owner isn't there

Agents often run while their owner is asleep or busy. The safe default:

1. Register, set goals and check in as usual.
2. Adopt only items with `owner_approval_needed: false`.
3. Queue the rest with the one-line reasons above.
4. Don't install, connect, pay or publish until the owner says yes, even if the agent technically could.

In Agent Coach the agent sets `"constraints": {"owner_present": false}` and the recommendations come back with that instruction attached ([setup doc](https://upskill.coach/start)).

## What approval is not

Approval is not a security check, and a security check is not approval. Our scan reads a pinned commit for prompt injection, malicious code, leaked secrets and vulnerable dependencies ([method](https://upskill.coach/docs/security-scanning), [findings across the library](https://upskill.coach/research/agent-skill-security)); the owner's yes is about authority. An item can pass the scan and still be wrong for this owner: it may need an account they don't want to create, cost money they didn't budget, or post under a name they don't want used. Ask for both.

## What we have not measured

These are 121 items from a curated catalog for founder, go-to-market and operator agents, not a random sample of public skills. The approval flag follows a fixed rule, not a judgment per owner; some owners will pre-approve whole classes (for example "any free MCP server from this vendor"). We have not measured how often agents actually wait for approval.

## Let a coach keep track

Agent Coach tells your agent, for every recommendation, whether it needs your approval and why, and checks in daily on whether each item earns its place. It is free. Browse the [library](https://upskill.coach/library), read the [developer docs](https://upskill.coach/docs), or paste one line into your agent:

```
Read https://upskill.coach/start and follow it.
```

Related: [how to vet an MCP server or skill](https://upskill.coach/blog/how-to-vet-an-mcp-server-or-agent-skill), [how an agent should store API keys](https://upskill.coach/blog/how-should-an-ai-agent-store-and-use-api-keys-for-mcp-servers), [remote or local MCP servers](https://upskill.coach/blog/remote-vs-local-mcp-servers-which-is-safer).

## FAQ

### Can an AI agent install a skill or MCP server without asking its owner?

It shouldn't, unless the item is plain text. Anything that runs code, needs a key or account, costs money or acts in public should wait for the owner's yes. In our catalog on 2026-11-19, 23 of 121 vetted items (19%) need no approval, and all of them are prompts or text-only checklists.

### What counts as needing owner approval?

Four things: it runs code on the owner's machine or account (a skill with scripts, an MCP server, a repo or library), it needs an API key or account, it costs money, or it sends or publishes something in public under the owner's name. One of these is enough.

### Is a SKILL.md file safe to adopt without approval?

Only if it is text you read and nothing else. A SKILL.md can ship scripts and tell the agent to run commands. In our catalog 6 skills are text-only checklists the agent can adopt alone; every other skill is marked as needing approval because it installs files, needs an account or acts in public.

### What should an agent do when its owner isn't around?

Set up and keep working, but only adopt items that need no approval. Queue the rest as a short list with what each one needs (code, key, cost, public action) and ask once when the owner is back, instead of asking item by item.

### Does a passing security scan remove the need for approval?

No. A scan checks one commit for known bad patterns; approval is about authority (whose machine, whose account, whose money). All 32 vetted MCP servers in our catalog passed review and every one still needs the owner's OK.

Questions or corrections: ops@avatar33.com.
