Use it when
- Your team repeats instructions, checklists or formats in every conversation.
- AI needs to read or update information in company services.
- You want to distribute a ready-made capability to other people or agents.
Do not start here when
- A clear one-off instruction already solves the task.
- The process changes every week and no one has defined the expected result.
- The connection would require more access than the job needs.
Start with the problem, not the acronym
Picture a sales team that gathers CRM data every Friday, checks stalled deals and writes an executive update. Repeating the same analysis instructions is a method problem. Copying CRM data into a chat is an access problem.
They look similar inside a conversation, but they need different components. Mixing them often creates expensive, brittle setups with unnecessary permissions.
Skill: a procedure AI can follow
A skill combines instructions and resources for repeatable work: steps, quality criteria, output templates and approved examples. It is useful once the company knows how the job should be done.
A proposal-review skill might check scope, deadline, margin, risks and next actions. It does not grant CRM access by itself.
Connector: access with identity and permission
A connector links AI to a service such as Google Drive, Slack, GitHub or a CRM. It carries identity, authentication and action boundaries. Its practical question is: what can this person read or do?
Use one when copying and pasting has become part of the process. Start with read-only access, a narrow data set and a test account.
MCP: the shared interface for AI tools
MCP is an open protocol for presenting tools and context to AI models. It is not the workflow or the business method. It lets compatible clients discover and call tools in a standard way.
It becomes valuable when one internal capability should work across multiple AI clients. An inventory MCP server, for example, could expose tools to check stock, locate an order and register an incident.
Plugin: the installable package
In OpenAI's ecosystem, plugins can bundle skills, connectors, MCP servers and other components. Think of the plugin as the package and the other pieces as capabilities inside it.
For a small internal workflow, a skill plus a connector may be enough. Packaging matters when you need a clear installation and configuration path for many users.
A practical decision rule
Write the outcome in one sentence first: “every Monday, find proposals with no reply and prepare a follow-up list.” If the missing piece is method, build a skill. If it is data or actions, assess a connector. If the capability must serve multiple AI clients, consider MCP. If it must be distributed as a product, consider a plugin.
Block uses MCP to enable actions without distributing API keys
In Anthropic's published case, Block connects its internal goose agent to Databricks and proprietary systems. MCP is used when AI needs controlled capabilities, while OAuth with short-lived credentials preserves each person's identity and access.
About 4,000 of Block's 10,000 employees used goose across 15 work profiles.
The case reports that 75% of engineers saved eight to ten or more hours per week.
An MCP server teaches the agent to submit code to Block's internal Beacon service.
The results come from a model vendor's customer story, not an independent audit.
MCP does not replace process, permission or quality criteria. It fits when AI needs a live, controlled capability. If the routine only needs repeatable guidance, a skill remains smaller, cheaper and easier to maintain.
Choose the right building block before creating an integration
Diagnose one routine. The output should be understandable to a business owner and reviewable by a technical owner.
1. Define the job
Record trigger, inputs, decisions, output and review owner.
An observable task that does not start with a technology name.
2. Test with context only
Use a prompt and authorized references before requesting system access.
Evidence of whether live data is actually required.
3. Turn the method into a skill
Package sequence, criteria, examples and boundaries when consistency is the problem.
A repeatable workflow with no server.
4. Look for an existing connector
Verify provenance, permissions, actions and data policy.
Less custom code and an explicit access list.
5. Justify a plugin or MCP
Build only when authentication, live data or controlled actions cannot be supplied safely by an existing connection.
A specification for small tools, schemas and authorization.
6. Test abuse and maintenance
Simulate unauthorized access, incomplete input, outages, revocation and version changes.
Blocking criteria, logs and a maintenance owner.
Choose between prompt, skill, connector, plugin and MCP
Use this diagnosis before opening an integration project. The complete pack in development will add a decision matrix, permission inventory, tests and server specification.
Help me choose the smallest architecture that can support this routine. ROUTINE [what happens today, who participates and which output must exist] SYSTEMS [files, CRM, email, calendar, database or others] REQUIRED ACTIONS [retrieve, summarize, draft, update, send or approve] CONSTRAINTS [sensitive data, permissions, audit, budget and deadline] Compare four options: 1. a well-structured prompt only; 2. a skill with instructions and references; 3. an existing connector or integration; 4. a custom plugin or MCP server. For each option, state what it solves, what it does not solve, required access, operational risk, maintenance and the signal that it has become too limited. Recommend the smallest viable option and design a reversible pilot. Do not recommend MCP merely because more than one system is involved.
The guide stays open. You pay for the shortcut.
Join the early list. We will use demand to decide which pack should be released first and what it must include.
- Architecture decision matrix
- Data and permission inventory
- MCP tool specification
- Authorization and failure tests
- Maintenance and shutdown plan
Verify at the source
Product capabilities and policies change. These are the official references reviewed for this guide.