Glossary

Model Context Protocol (MCP)

One protocol between models and the systems they need, instead of one integration per pair.

Definition

The Model Context Protocol (MCP) is an open standard, released by Anthropic in November 2024, for connecting AI applications to external tools and data. An MCP server exposes tools, resources and prompts over a defined JSON-RPC interface; an MCP client inside an AI application discovers and calls them. One protocol replaces bespoke per-integration code, so any compliant client can use any compliant server.

Released: November 2024
Published by: Anthropic
Licence: Open standard, open source
Transport: JSON-RPC

The problem it was designed to solve

Before a shared protocol, connecting M AI applications to N data sources meant writing M times N integrations, each with its own authentication, schema and failure behaviour. Every new model host reimplemented the same connectors, and every tool vendor reimplemented the same adapters for each host. MCP turns that into an M plus N problem: implement the protocol once on either side and the combinations come for free.

Anthropic published the specification and reference implementations in November 2024 and open-sourced the surrounding SDKs. Adoption spread beyond Anthropic's own products through 2025, and by 2026 MCP is the most widely implemented interface for exposing tools and context to language models.

The three things an MCP server can expose

The specification defines distinct primitives, and mixing them up is the most common implementation mistake.

Tools
Functions the model can invoke, with a described input schema. Model-controlled: the model decides when to call them, subject to whatever approval the host enforces.
Resources
Data the client can read and attach to context, such as files, records and query results, identified by URI. Application-controlled rather than model-invoked.
Prompts
Reusable templates a user can select, typically surfaced in the host application as slash commands or menu entries. User-controlled.

Commonly confused with

Term	What it is	The difference
Function calling	A model capability: emitting a structured call against a supplied schema	MCP standardises how tools are discovered, transported and authorised. Function calling is what happens once a tool is in front of the model.
Plugin API	A vendor-specific extension interface for one product	Tied to one host. MCP is host-neutral by design.
Agent framework	A library for building agent loops, memory and orchestration	MCP does not run agents. It is a connection layer that agents use.
API gateway	Infrastructure routing and securing HTTP APIs	Serves general clients. MCP describes capabilities in terms a model can reason about.

A concrete example

A team wants an assistant to answer questions from their Postgres database. Without MCP, someone writes a bespoke connector for whichever assistant they use, and writes it again when they switch. With MCP, they run a Postgres MCP server that exposes a query tool and the schema as resources. Any MCP-capable client can then be pointed at it, and the same server also serves the coding agent and the internal chat app.

Polaris does not require MCP to give an AI worker tool access. Connections come from a curated catalog, are authorised once for the whole organisation, and are stored server-side so workers can use credentials that browsers can never read back.

Skill file

The capability an agent has, written down where a person can read and edit it.

Agent runtime

Not the model, not the framework: the thing that actually runs the job.

Autonomous agent

Autonomy is a range, and the interesting question is where the boundary sits.

Every tool a Polaris worker can be given

One catalog, one credential per tool per organisation, authorized by an owner and used by every worker who carries it.

How an AI worker gets tool access

The interesting part of tool access is not the list. It is where the credential lives and who can read it.

Agent orchestration

Deciding which agent does what, in what order, and what happens when one fails.

Questions people ask

+Who controls the Model Context Protocol?

Anthropic published the specification in November 2024 and maintains it as an open standard with open-source SDKs and reference servers. It is not proprietary to Anthropic's products: other model providers, IDEs and agent hosts implement the same protocol.

+Is MCP a security risk?

It is a connection layer, so it inherits the risk of whatever it connects. The specific concerns are prompt injection through resource content, over-broad tool permissions, and servers from untrusted sources. The mitigations are conventional: least-privilege credentials, human approval for consequential tool calls, and running only servers you have reason to trust.

+Do I need MCP to give an AI agent access to my tools?

No. MCP is one way to expose tools to a model, and direct API integrations remain common, particularly where a product curates its own connection catalog. MCP's advantage is portability: the same server works with any compliant client rather than one vendor's.

+What does JSON-RPC have to do with it?

JSON-RPC is the message format MCP uses for requests and responses between client and server. It matters to implementers rather than users, but it is why an MCP server can run locally over standard input and output or remotely over HTTP without changing its logic.