MCP vs ACP vs AI skills: choose the right connection
Use concrete examples to distinguish model endpoints, MCP tools, ACP coding agents, reusable skills and scheduled tasks.
You want an assistant to read an issue, fix a bug or repeat a weekly check. “Add an integration” doesn’t tell you which setup you need. Model endpoints, MCP, ACP, skills and schedules each handle a different part of the job.
Start with the action you want and work out what’s missing. Adding a protocol doesn’t give you a better model, sign you in to an account or keep a computer switched on. This guide helps you pick the smallest setup that works, in Mono by Reunified or a similar app.
1. Identify the missing capability
Use this table before you go looking for a plugin:
| Your immediate goal | Start by checking |
|---|---|
| Use a different model for text replies | Model endpoint settings |
| Read or act on information in another service | A suitable MCP server or built-in integration |
| Work with a coding agent on a repository | Whether the ACP client and agent work together |
| Reuse a packaged workflow | A skill and what it needs |
| Repeat an action that already works | A task scheduler and where it runs |
These are starting points, and they often work together. A skill can use MCP, a coding agent can call tools, and a scheduled task can run a skill. Find the first missing piece instead of installing every layer.
2. Use MCP for a service capability
Model Context Protocol (MCP) connects an AI app to features that a server offers. The official architecture guide splits them into tools, resources and prompts. A server may offer one or several of them.
Say you want to list open issues from a project service. You need a connection that offers that lookup, and an account that is allowed to read the project. A server that can search issues can’t necessarily create or close them.
Check what the server can actually do, and test it with a small read-only request. For setup steps, use the MCP connection guide. Switching to a different language model won’t give it access to your project account.
3. Use ACP for a coding-agent session
Agent Client Protocol (ACP) defines how a coding agent talks to a client, such as an editor. The ACP introduction explains this. The agent does the work in the repository; the client is the screen you use to direct it.
For “find and fix this failing test”, check that the agent, the client and the computer that runs them work together. Supporting the same protocol doesn’t prove they’ve been tested together. The repository and commands still need a computer with the right access and tools installed.
A phone can control that work without being the computer that runs it. In Mono Code, your computer hosts the projects and your phone drives agents such as Claude Code, Gemini CLI or OpenCode over ACP. Our remote coding guide covers small tasks, approvals and reconnecting. ACP doesn’t give you an always-on cloud computer, and it doesn’t pay the agent provider’s bill.
4. Use a skill for reusable behavior
“Skill” means different things in different products. In Mono, a skill is an add-on that can include modules, parameters and connections to other services. Elsewhere, a skill may be a document of instructions or a different package format. Don’t assume one product’s skills will work in another.
For “format these notes the same way every week”, first check what a skill expects to receive and what it returns. It might use only the text you give it, call a model, or fetch information through a connection. Each of these adds setup and one more thing that can fail.
Use the skill installation checklist to review what a skill needs and test one result. A successful install doesn’t prove that its data source works today or that it runs on every device.
5. Add scheduling only when the action works
A scheduler decides when work starts. It doesn’t supply missing inputs or permissions. For a weekly project summary, first check by hand that you can fetch the right items and get a useful summary. Then decide where the job can run at the time you need.
Mono Tasks run on your device, and a task missed while the app was closed catches up the next time you open it. So don’t expect a task to run on time while your phone is off. A task that needs a desktop command also needs that desktop; a schedule doesn’t move the command onto your phone.
When a scheduled result is missing, check where the task runs before you change the model. When a service returns an authorization error, check its account connection before you reinstall the coding agent.
6. Draw your own five-line setup
Write down five lines: model, outside services, reusable actions, coding computer (if any) and schedule (if any). Leave unused lines empty. Next to each part you use, note who runs it, what information it receives and who bills you for it.
A simple writing task may need only a model. Fixing a repository may need an ACP agent and a computer, but no new MCP connection. A weekly digest from a service may need a connection, a skill and a scheduler that is running.
Mono isn’t out yet, but you can join the waitlist. Meanwhile, use this map to plan a setup you can explain and test before you make it bigger.
