GrantbySince.dev
Sign inStart free

Guide · Updated October 8, 2026

How to connect Claude or ChatGPT to an MCP server

MCP (Model Context Protocol) is an open standard that lets an AI app call tools in other systems. To connect Claude or ChatGPT to your ops data, you add a remote server by its URL, sign in with OAuth or an API key, and start read-only.

On this page

What MCP is

MCP stands for Model Context Protocol. A vendor builds one MCP server, and any AI app that speaks MCP can use it, so nobody writes a custom integration for each app.

Anthropic released it in November 2024. In December 2025 Anthropic handed it to the Agentic AI Foundation, and by then there were more than 10,000 active public servers.

You'll see three words in the docs. The host is the app you chat in, like Claude, ChatGPT or Cursor. The server is the other end, and it offers tools such as listing failed CI runs or reading this month's AWS cost. The client is the connection between them, one per server. Servers can also offer resources (data to read) and prompts (templates), but ops servers are mostly tools (architecture page).

Local or remote MCP servers

To use a server in Claude or ChatGPT, you need a remote one, which means a URL. ChatGPT only connects to remote servers, and one on a private network needs a Secure MCP Tunnel (help article). Claude Desktop can also run a local server as a desktop extension (Claude docs).

Where an MCP server can run and what each kind suits
KindWhere it runsWhat it suits
Local, stdioYour machineFiles and tools on your own computer
Remote, Streamable HTTPThe vendor's platformHosted services such as GitHub, Sentry or Since

If a vendor only offers the older HTTP+SSE transport, ask when they're moving. Claude and Claude Code mark it deprecated (Claude, Claude Code).

We'd pick a vendor's hosted server over a community package you run yourself, because a bad start command can run any code on your machine (MCP security page).

MCP OAuth or API key

Our take: use OAuth for people in chat apps, and one API key per script, CI job or coding agent. Then you can revoke one without breaking the rest.

Whatever you use, the token goes in the Authorization header, never in the URL (authorization spec). The spec describes OAuth for HTTP servers, but many servers take an API key instead (architecture page).

With OAuth, each person signs in as themselves, and you can revoke their access later. In Claude's connector dialog, pick the option called Use Claude's published identity, which Claude's docs recommend (Claude docs). In Claude Code, run /mcp, select the server and choose Authenticate (quickstart).

An API key is one fixed credential. In Claude, choose No sign-in and paste the key under Request headers. That option is a beta, and only some organizations have it (Claude docs). On Team and Enterprise, everyone then shares one key, so everyone gets the same access (Help Center).

How to add a remote MCP server

Where to add a remote MCP server in Claude and ChatGPT
AppWhere to add itWho can
Claude on the web or DesktopCustomize, Connectors, Add custom connectorFree, Pro and Max users. Free can add one. On Team and Enterprise, an Owner adds it under Organization settings, Connectors, Add, Custom
ChatGPT on the webDeveloper mode, then Apps, Create, Scan ToolsFull MCP is a beta for Business, Enterprise and Edu workspaces

Menu names change. If one's missing, check the vendor's page.

In Claude, paste the server URL under Add custom connector. You can't edit sign-in settings later, so if they're wrong, remove the connector and add it again (Claude docs).

In ChatGPT, enter the endpoint and how it authenticates, then click Scan Tools. The app stays a draft until an admin publishes it (help article). OpenAI's developer docs show a second path: open ChatGPT Plugins, select the plus button, then Add custom MCP server.

In Claude Code, a project's servers go in .mcp.json. A remote entry needs its type set to http, or Claude Code treats it as a local server and fails (Claude Code docs).

Cursor, VS Code and Codex each read their own config file:

Before you give a model a new server, run npx @modelcontextprotocol/inspector, add the URL and read the tool list (Inspector docs).

Connect Claude to GitHub

This works in Claude Code. GitHub says its hosted server isn't supported in Claude Desktop and points Desktop users to a local Docker setup instead (GitHub), so we'd use Claude Code.

Add /readonly to the URL for read-only mode (GitHub docs), and use a fine-grained token that only covers the repos you need (token docs).

{
  "mcpServers": {
    "github": {
      "type": "http",
      "url": "https://api.githubcopilot.com/mcp/readonly",
      "headers": {
        "Authorization": "Bearer ${GITHUB_PAT}"
      }
    }
  }
}

Keep GITHUB_PAT in your environment, not the repo. Claude's dialog has a Use your own OAuth client option (Claude docs), but we haven't tested it with GitHub.

MCP security basics

Start read-only

Give the model read access first, and add write access only when it needs it. MCP's own advice says the same (MCP security page). In Claude, block the tools you don't need (Claude docs). For AWS, use a read-only IAM role. The AWS Labs servers for CloudWatch, CloudTrail, and Billing and Cost Management don't mention a read-only setting, so enforce it with the role (AWS Labs).

Watch for prompt injection

An issue, a log line or a web page can carry instructions for the model. OWASP calls this indirect prompt injection.

Simon Willison calls the dangerous mix the lethal trifecta: private data, content an attacker can write, and a way to send data out. In ops, that's a model that reads your private repos, reads public issues anyone can open, and can post comments. Don't give one connection all three. If it reads public issues, don't let it post on its own. OpenAI and Claude Code both warn about prompt injection through MCP servers (OpenAI, Claude Code).

Approve writes yourself

When a vendor ships an update, re-read its tool list before you approve anything new. The spec says a person should always be able to deny a tool call (tools spec). In Claude, only use Always allow on servers you trust, and turn off write tools in Research mode (Help Center). ChatGPT may ask before write actions, and it freezes an approved app's tool list until an admin reviews changes (help article).

Ops questions MCP is good for

Start with questions that only read, like these:

  • What changed since last week?
  • Why did CI fail on main?
  • Which production errors are new?
  • What did the AWS account cost this week, and what moved?
  • Which pull requests have waited longest?

GitHub's server covers repositories, issues and pull requests. Sentry runs an official remote server with OAuth that can be scoped to one project (Sentry). For cost, use the AWS Labs servers.

A connected model can still be wrong. Ask it for IDs and links, and click through a few. We'd connect two or three servers to start. These questions fit a weekly operations review. For cost, see AWS cost monitoring, and for alerts, see security alert triage.

Since's MCP server, and what Grant doesn't do

Grant, COO

Since has a remote MCP server, so you can ask Grant, who leads Since's team of AI agents, from your own AI app. The team works from the sources you connect, a supported fix arrives as a pull request for you to review, and nothing merges or deploys without a person.

Since's remote server is at https://api.since.dev/mcp. It has 21 tools: 6 for Grant, and 15 platform tools for the outside changes Since watches, your repositories and your usage (MCP reference).

In Claude, Codex and Cursor, paste the URL and sign in through the browser. Anywhere else, make a workspace API key under Connections, then External agents, and send it as a Bearer token, like this:

{
  "mcpServers": {
    "since": {
      "type": "http",
      "url": "https://api.since.dev/mcp",
      "headers": {
        "Authorization": "Bearer ${SINCE_API_KEY}"
      }
    }
  }
}

Any member can make a key, but a new key has no Grant access. An owner or admin picks the projects it can see and its allowed actions: read, question, investigate, decision and control. Browser sign-in asks for the same choices. We'd start with read alone, which covers the first bullet below.

The Grant tools need Pro, Studio or Enterprise. On Free they return a plan_required error, and the platform tools still work within your plan's limits. Pro is $49 a month (see plans).

Once you're connected, start with these:

  • grant_team_status shows your approved projects, permissions and Grant's team, and grant_read reads saved findings, runs and alerts. Neither calls a model, so neither uses credits.
  • grant_question (needs question, plus investigate for live team answers) asks Grant about your approved projects and waits for the answer (tool reference).

Over MCP, Grant can't write to your cloud, rerun CI, merge or deploy (authority table).

A question over MCP uses Since AI credits just like one in the app. If "Use my own AI provider" is on, it runs on your own provider and uses none (grant_question docs).

Questions

What is an MCP server?

An MCP server is a program that gives an AI app tools and data from another system.

What is the difference between a local and a remote MCP server?

A local server runs on your machine. A remote one is hosted, and you connect by URL. Claude on the web and ChatGPT need a remote URL.

Do I need an API key or OAuth to connect to an MCP server?

Either, depending on the server. We'd use OAuth when a person connects from a chat app, so each person signs in as themselves. Use an API key for scripts, CI jobs and coding agents, one key each.

Can ChatGPT use MCP servers, or only Claude?

ChatGPT can. On the web, full MCP support is a beta for Business, Enterprise and Edu workspaces, and ChatGPT Pro can connect servers with read and fetch permissions in developer mode (help article).

How do I connect Claude to GitHub with MCP?

In Claude Code, add GitHub's hosted server, https://api.githubcopilot.com/mcp/readonly, to .mcp.json with type http and a fine-grained token in an Authorization header. GitHub says the hosted server isn't supported in Claude Desktop and suggests its local Docker setup there (GitHub).

Is it safe to connect an MCP server to my company's systems?

It can be. Start read-only, give each agent its own limited credential, keep a person approving writes, and only connect vendors you trust.

Which Since plans include the MCP server?

All of them. The platform tools work on every plan. The Grant tools need Pro, Studio or Enterprise, and on Free they return a plan_required error instead of an answer.