A local model by itself only sees the prompt you give it. It cannot read your files, inspect GitHub issues, search a database, or check a calendar unless something connects it to those sources.

MCP, the Model Context Protocol, is one way to standardize that connection.

Mental model: MCP is a tool plug

Anthropic’s MCP documentation describes MCP as an open protocol that standardizes how applications provide context to LLMs. The common analogy is a standard connector between AI applications and data or tools.

In beginner terms:

PieceMeaning
MCP clientThe app or agent that wants tools and context
MCP serverA process that exposes tools, resources, or prompts
ToolAn action the model can request, such as search, fetch, create, or update
ResourceReadable context, such as a file, document, issue, or database entry
Permission layerRules about what the model is allowed to access or change

Model Context Protocol (MCP)

Click a component to see its role

LLM
Generates Text / Tool Calls
MCP Client
Your App / Agent
Filesystem Server
Reads local files
GitHub Server
Reads Issues/PRs
The model does not become all-powerful. The client decides what servers are connected and what actions require approval.

Official source to verify:

Why MCP matters for local LLMs

Local LLMs are often attractive because of privacy and control. But isolated local chat is limited. MCP can help local or hybrid tools connect to controlled context sources.

Examples:

MCP server ideaWhat the model can do
FilesystemRead selected project files or notes
GitHubInspect issues, pull requests, comments, and files
Browser/searchFetch current pages or search documentation
DatabaseQuery structured project or research data
Notes appRetrieve personal knowledge
Local evaluation harnessRun fixed prompts and return scores

The key word is selected. A safe MCP setup should expose only what the task needs.

Tools vs resources

Tools and resources are easy to mix up.

MCP conceptBeginner example
Resource“Here is the content of README.md.”
Tool“Run the test command and return the output.”
Prompt“Use this reusable prompt template for a task.”

Resources provide context. Tools perform actions. Tools usually need more caution because they can change state or access external systems.

Local model plus MCP

A local model can participate in an MCP workflow when the application around it supports MCP or bridges MCP calls into the model loop.

A simplified loop:

  1. User asks a question.
  2. Agent app sends prompt and available tool descriptions to the model.
  3. Model asks to call a tool.
  4. MCP client calls the MCP server.
  5. Server returns data or action results.
  6. App sends the result back to the model.
  7. Model answers or asks for another tool call.

The model is not directly browsing your computer. The app mediates every step.

Key papers on tool use, function calling, and retrieval-augmented generation

OpenAI-compatible local endpoints

Some local runtimes expose OpenAI-compatible APIs. Ollama documents support for OpenAI-compatible chat completions fields including messages, streaming, JSON mode, tools, and tool choice. This compatibility can make local runtimes easier to connect to existing apps.

But OpenAI-compatible does not automatically mean fully equivalent. Check exact support for:

FeatureWhy to verify
Tool callingSome models produce tool calls more reliably than others
Structured outputsSchema enforcement may differ
VisionImage input depends on model and runtime
StreamingClient and server must both support it
Context lengthRuntime settings may limit effective context

Official docs to verify:

Permissions matter

MCP can connect useful tools, but it also expands the blast radius of mistakes. A local LLM with access to a shell, browser, filesystem, and GitHub is no longer “just local chat.”

Beginner safety checklist:

  1. Start read-only when possible.
  2. Use project-scoped filesystem access.
  3. Require confirmation for writes, deletes, shell commands, and network actions.
  4. Keep API keys outside prompts and logs.
  5. Prefer narrow tools over broad shell access.
  6. Review generated changes before applying them.

What to verify because tools change

Verify MCP support in the specific client you use. Claude Code, Codex, desktop apps, IDEs, and local agent frameworks may expose MCP differently. Also verify whether a server is official, community-maintained, or custom.

The practical lesson: MCP is not the model. It is plumbing for context and tools. For local LLMs, that plumbing can turn a private chat model into a useful assistant, but permissions decide whether it is safe.