You connect an agent to GitHub. It can read issues and inspect pull requests. But it still does not know what your team considers a good review, which checks matter, or how you want the findings written.
That is the gap between access and a process. MCP helps with the access. A skill can describe the process. Most useful agents need both.
Skills vs MCP at a glance
| Question | Agent Skills | MCP |
|---|---|---|
| What do you package? | Instructions and supporting material for a task | A standard interface between an AI application and a server |
| Typical contents | A SKILL.md file, examples, references, optional scripts | Tools, resources, and prompt templates |
| What does it help answer? | How should this work be done? | What can this application access or call? |
| Example | Our pull-request review procedure | Access to repository issues and pull requests |
| What still needs setup? | The tools and permissions the procedure requires | Authentication, supported client features, and the workflow around the connection |
Skills can include scripts; MCP servers can expose prompts as well as tools. See the Agent Skills overview and MCP introduction for the two formats.
What is an agent skill?
An agent skill is a reusable package of instructions and resources. In
the open Agent Skills format, its entry point is a
SKILL.md file. A host can discover a skill from its name
and description, then load more of the package when it is relevant. This
is called progressive disclosure. The
format specification
defines the metadata and directory conventions.
For a weekly update, your skill might explain which projects to cover, how to distinguish completed work from plans, and what a good final update looks like. A reference file could hold your team's writing examples. A script could format the finished data into a document.
What is MCP?
The Model Context Protocol is a standard for connecting AI applications to external systems. A client connects to a server and uses the capabilities that the server exposes and the client supports. The protocol gives integration authors a shared interface instead of requiring a different bespoke integration for every host. See the official introduction.
MCP servers can expose three particularly useful primitives: tools that perform operations, resources that supply context, and prompts that provide reusable templates. A server does not have to implement all three. A host also needs to support the particular feature you intend to use. The server concepts guide explains these differences.
For concrete options, our best MCP servers guide groups servers by the work you want to do.
A practical example: preparing a release update
Suppose you ask an agent to prepare this week's release update from GitHub and Linear. There are several separate jobs hiding inside that request.
An MCP connection can provide the operations for retrieving issues and pull requests. A skill can tell the agent to include merged user-facing changes, separate fixes from new features, link each claim to its source, and flag anything it cannot verify.
A compact instruction file might look like this:
---
name: release-update
description: Draft a weekly release update from completed project work.
---
1. Confirm the project and date range.
2. Find completed work using the available project connections.
3. Verify each item against its source before including it.
4. Group the draft into Features, Fixes, and Known limitations.
5. Add a source link to every item.
6. Return the draft for review before publishing it.
To use this example, connect the project services and give the agent access to the relevant repositories and issues.
How this works in Dexto
Dexto's Skills Library is where you browse and manage skills for a desktop. You can inspect a skill's instructions and files, see installed skills in My Library, and find Dexto and community skills in the Store. Dexto loads relevant skill information as it works on a task.
Connections provide access to services such as GitHub, Linear, and Notion. Dexto discovers the tools offered by a connected service and also supports custom MCP servers. For the release-update example, you would connect the project services, install or create a suitable skill, and ask for the update.
MCP also works in the other direction: Dexto exposes an MCP server so you can access Dexto from clients such as Claude, Codex, and Cursor. The MCP setup guide covers that connection. Connecting an external client to Dexto and connecting a service inside Dexto are different setup steps.
If the workflow includes running scripts or producing files, it may also need an execution environment. Our AI sandbox platforms comparison explains that part of the stack.
Which should you add first?
Start with the failure you are trying to fix.
If the agent cannot reach the information, add the connection. If it can reach the information but keeps following the wrong process, write or improve the skill. If it needs to run a program, make sure an appropriate execution tool is available too.
For example, “review this spreadsheet using our finance template” needs a defined review method and access to the spreadsheet. “Find the latest open issues” primarily needs a working project connection. “Generate a chart from these issues every week” adds scheduling, code execution, and a place to keep the output.
This is why the agent harness matters. It coordinates the model, tools, context, execution, and approvals around the task.
Do skills replace MCP servers?
A skill can use a CLI or a direct API through an available execution tool, so MCP is not required for every integration. But the skill still needs a way to execute that operation and authenticate it. Writing an instruction to call an API does not create the connection.