What is an MCP server? And how to point one at your issue tracker

An MCP server is a program that exposes a set of tools to an AI agent over a standard protocol, so the agent can take actions and read data outside its own context window. MCP stands for Model Context Protocol, an open standard introduced by Anthropic in late 2024. The agent is the client; the MCP server is whatever sits on the other end, whether that is a database, a file system, GitHub, or an issue tracker. Connect a server once, and any agent that speaks MCP can use its tools.

That is the whole idea. The rest of this post is the part the definition skips: what an MCP server actually exposes, how it differs from a plain API, and what changes when the server on the other end is your issue tracker.

#What an MCP server exposes

An MCP server publishes three kinds of things, and an agent discovers all of them automatically when it connects:

• Tools. Actions the agent can take: run a query, send a request, create a record. These are the verbs.
• Resources. Read-only data the agent can pull in: documents, logs, a codebase, a list of records.
• Prompts. Pre-defined templates for common, multi-step tasks the server wants to make easy.

For an issue tracker, the tools are the verbs you already use: create an issue, search issues, list issues, comment, close an issue, run the triage queue. When the server is connected, those tools show up in the agent's tool list next to its built-in file and shell tools, and the agent calls them the same way. "What is left on RAD-219" stops being a thing you go look up and paste back, and becomes a thing the agent does in its own loop.

#MCP server vs API: what is the difference?

This is the most common follow-up, so it is worth being precise. A plain API is a fixed set of endpoints that another program calls. You, the developer, read the docs, write the integration code, and maintain it as the API changes. The API does not describe itself to a model; you describe it, by hand, every time.

An MCP server wraps that work in a standard the agent understands. The server advertises its tools, their inputs, and their outputs in a format the agent reads at connection time. So the agent discovers what it can do without you writing a custom connector for each service. The common analogy is USB-C for AI: one shape of plug, and anything that speaks it works without a new cable per device.

In practice an MCP server usually sits in front of an API. Radial, for example, has a REST API for programs and an MCP server for agents, and they expose the same underlying tracker. The REST API is what you script against directly; the MCP server is what an agent connects to so it can decide which call to make on its own.

#How to point an MCP server at your issue tracker

The reason this matters for a tracker specifically: the tracker is where work is decided, and an agent that can read and write it stops needing you as the message bus. Radial ships a hosted MCP server at mcp.radial.build and serves the same tools locally over stdio through its CLI. The local path is one command:

radial mcp

That starts the MCP server over stdio, exposing the eight tracker tools (create_issue, update_issue, search_issues, list_issues, comment, close_issue, list_projects, triage_queue) to any MCP client you point at it. For a hosted connection from Claude Code instead, you add the remote server and authorize it in the browser, with no key to paste. The deeper walk-through of that flow is in Give Claude Code access to your issue tracker in 5 minutes.

Once the server is connected, a session reads like plain instructions: search for the open issues assigned to me, start on the highest priority one, close it with a comment when it is done, file a new issue for the second bug you found along the way. The agent reads the work, does the work, and writes the result back to the same place. That place is also where the next agent session will look, which is the quiet payoff: the reasoning behind a change lives in the issue the agent already filed, not in a chat window that resets when it fills up.

#FAQ

#What is the difference between an API and an MCP server?

An API is a fixed set of endpoints another program calls, and you write the integration code for each one by hand. An MCP server wraps those capabilities in a standard the agent reads at connection time, so the agent discovers and uses the tools without a custom connector per service. An MCP server often sits in front of an API; it is the layer that makes the API self-describing to a model.

#What is an MCP server used for?

Giving an AI agent the ability to act and read outside its own context: query a database, search a codebase, file an issue, send a request. Anything you would otherwise hand-integrate, exposed as tools the agent can call directly.

#Does ChatGPT use MCP?

MCP is an open standard, not specific to one vendor, and support has spread across the major agent clients since Anthropic published it. The protocol is deliberately client-agnostic, so a server you build works with any client that speaks MCP rather than being tied to a single product.

#Is MCP like TCP?

By analogy, yes: both are standards that let independent systems talk without a custom integration for each pair. TCP standardizes communication between machines on a network; MCP standardizes how an AI agent reaches external tools and data. The comparison is about the role of a shared protocol, not the technical layer.

#How do I create an MCP server?

You implement the protocol's tool, resource, and prompt interfaces and serve them over a transport (stdio for a local process, HTTP for a hosted one). Most trackers and platforms that already have an API are now shipping an MCP server in front of it so agents can connect without custom code. With Radial you do not build one for the tracker; radial mcp serves it locally and mcp.radial.build hosts it.

#Does connecting an agent over MCP cost extra?

Not with Radial. Agent credentials never count as billed seats. You pay $50 per user, per year, for the humans on the workspace, flat and billed annually, locked at the rate you join. There is no AI credit meter anywhere in the product, and under the Plain Software Pledge, the day we ship a copilot, meter usage, or charge for AI you did not ask for, your subscription is free.

#The short version

An MCP server is the standard bridge between an agent and the tools and data it needs. Point one at your issue tracker and the agent reads the work, does the work, and writes the result back where it already lives. See the full tool list and the OAuth flow on the developers page.