Model Context Protocol (MCP)

The Model Context Protocol (MCP) is an open standard that defines how AI applications connect to external tools, data, and services. An MCP server exposes tools, resources, and prompts over JSON-RPC, and any MCP client, such as a coding agent, can discover and call them without a custom integration for each pair. Anthropic introduced MCP in November 2024, and it is now a project of the Agentic AI Foundation under the Linux Foundation.

What is the Model Context Protocol?

The Model Context Protocol, or MCP, is an open standard for connecting AI applications to the systems they need to act on. Before MCP, every pairing of an agent and a tool needed its own integration. MCP replaces that with one protocol: a service implements an MCP server once, and every MCP client can use it.
The design borrows from the Language Server Protocol, which did the same for programming languages and code editors. Anthropic released MCP in November 2024 and in December 2025 donated it to the Agentic AI Foundation, a directed fund of the Linux Foundation co-founded by Anthropic, Block, and OpenAI.
MCP is easy to confuse with MCAP, the container format for robot logs. The names differ by one letter, and a physical AI stack often uses both: MCAP holds the recorded data, and MCP is how an agent reaches the tools that work with it.

Key takeaways

  • MCP is an open protocol that standardizes how AI agents call external tools and read external data.
  • An MCP server exposes tools, resources, and prompts, and any MCP client can discover and use them.
  • What an agent can do through MCP is capped by the server, so permissions belong on the server side, and the host still asks you before a tool runs.

How it works

MCP has three roles. A host is the AI application, such as Claude Code, Cursor, or VS Code. Inside the host, each client talks to exactly one server. A server is a service that exposes capabilities. Messages use JSON-RPC 2.0, and a server runs either locally on the same machine as the host or remotely over HTTP, where sign-in typically uses OAuth.
Each client in the host talks to one server, and the server decides what an agent can do.
Servers offer three kinds of capability:
  • Tools are functions the model can call, such as running a query or launching a job.
  • Resources are data the application can read, such as files or records.
  • Prompts are reusable templates for common workflows.
When an agent connects, it asks the server what it offers, then decides which tools to call as it works. The specification treats every tool as arbitrary code execution, so hosts are expected to get the user's consent before invoking one.

Point an agent at real robot data, then check what it did

An agent's answer is only as good as the tools it reached for and the data behind them. When an agent works over MCP, check:
  • which operator it called, and whether that matches what you asked for
  • whether it ran on the dataset or view you meant, not a near match
  • what changed in the dataset, which the activity history records under your name
In Voxel51, the DROID dataset holds 1,000 real-world manipulation episodes with task prompts, grasp counts, and motion stats that an agent can query.

Robotics Droid

1,000 episodes

Real-world robot manipulation episodes from DROID, stored as MCAP with three synced cameras and fields for task prompt, success, and number of grasps.

Try this: Ask the Voxel51 Agent for the episodes with two or more grasps, then check which field it filtered on before you trust the list.

Explore DROID in Voxel51

Free with a work or .edu email.

Go deeper

Why it matters

Data work is where agents most need real tools. An agent that can only describe a curation step is guessing, while one connected to your datasets over MCP can run the step and show you the result. The Voxel51 Agent Gateway is a remote MCP server built for this. It connects Claude Code, Cursor, VS Code, GitHub Copilot, and other MCP clients to your own Voxel51 deployment, in the cloud or self-hosted. Its operators come from your installed plugins, and every call runs under the approving user's own role and dataset permissions and appears in the dataset's activity history.
The Agent Gateway connects any MCP client to your own Voxel51 deployment.
Inside the app, the Voxel51 Agent uses the open source FiftyOne MCP server, which exposes more than 80 operators for dataset management, model inference, and brain computations. Agent skills sit on top of MCP: the server provides what an agent can do, and a skill teaches it how and when to do it.
MCP connects agents to tools. MCAP records robot data.

Frequently asked questions

What is the difference between MCP and an API?

An API is specific to one service, so an agent needs custom code for each one. MCP is a common protocol on top of those services. A server describes its tools in a standard way, so any MCP client can discover and call them without new integration code.

What is an MCP server?

An MCP server is a program that exposes tools, resources, and prompts to AI applications over the Model Context Protocol. It can run locally on your machine or remotely as a hosted service.

Is MCP the same as MCAP?

No. MCP is a protocol for connecting AI agents to tools. MCAP is a file format for recording timestamped, multimodal robot data. They often appear together in physical AI workflows, which is where the mix-up starts.

Can I connect my own coding agent to Voxel51?

Yes. The Voxel51 Agent Gateway is a remote MCP server for Voxel51 Cloud and self-hosted deployments. Connecting takes one command or one prompt plus a browser approval, and self-hosted deployments need remote agent access turned on first.
black and white photo of Jesse Mostipak
Director of Growth
Jesse Mostipak is the Director of Growth at Voxel51, where the work is helping humans find and trust what the brand knows, and teaching the Google knowledge graph and the LLMs answering on their behalf to do the same. That question, how knowledge gets built inside a system, is one Jesse has been chasing for years. Earlier versions of it ran through a New York City high school science classroom, data science and machine learning, and developer relations at Kaggle, Posit (formerly RStudio), and Baseten. The answer doesn't change much depending on whether the learner is a teenager, a software engineer, or a knowledge graph. Jesse holds a Master's in Education from CUNY Hunter College. LinkedIn
See all articles by Jesse Mostipak
Last updated November 19, 2026

Building visual or physical AI?

Let's talk.