HostTracker

HostTracker MCP server

Connect an AI assistant to HostTracker over the Model Context Protocol and let it *operate* your monitoring account in conversation: run a live check from 300+ global locations, see what is down, create or pause a monitor, schedule a maintenance window, review incidents, manage who gets alerted, wire a webhook, publish a status page update.

Endpoint   https://mcp.host-tracker.com/mcp

Transport streamable HTTP Auth OAuth 2.1 (sign in when your client asks) - or Authorization: Bearer

This repository is the public face of that hosted server: the connection metadata (server.json), the per-client setup guide (CLIENT.md), and the security policy (SECURITY.md) - and, since 2.0.0, the server's own source code in src/, so you can read exactly what runs behind the endpoint or run a copy yourself (see "Run it yourself" below).

Connect in two minutes

With OAuth (recommended - Claude.ai, Claude Desktop, Claude Code, ChatGPT and any client with an OAuth-capable connector dialog):

1. Add the endpoint https://mcp.host-tracker.com/mcp as a connector in your client.
2. Sign in and approve. The client opens HostTracker's sign-in page, then a consent card listing the permissions it asks for (by default: run checks + read monitors). Press Approve. No token ever appears.
3. Ask in plain language: *"is example.com up right now, checked from Europe and Asia?"*, *"which of my monitors are down?"*, *"pause the staging monitor until tomorrow"*.

Claude Code

With OAuth (no token needed):

claude mcp add --transport http hosttracker https://mcp.host-tracker.com/mcp

then, inside Claude Code: /mcp -> hosttracker -> Authenticate (opens the sign-in + consent page)

Or with a bearer token in the header:

.mcp.json in your project (or ~/.claude.json for a user-wide connector):

{

"mcpServers": { "hosttracker": { "type": "http", "url": "https://mcp.host-tracker.com/mcp", "headers": { "Authorization": "Bearer YOUR_HOSTTRACKER_API_TOKEN" } } } }

Or from the command line:

claude mcp add --transport http hosttracker https://mcp.host-tracker.com/mcp \

--header "Authorization: Bearer YOUR_HOSTTRACKER_API_TOKEN"

Claude.ai and Claude Desktop

Settings -> Connectors -> Add custom connector -> paste https://mcp.host-tracker.com/mcp -> Connect. The browser opens HostTracker's sign-in + consent page; press Approve and the connector is live. That is the whole setup.

To use a bearer token instead (for example a long-lived token with a hand-picked scope set), Desktop can also connect through the mcp-remote bridge. Edit claude_desktop_config.json:

{

"mcpServers": { "hosttracker": { "command": "npx", "args": [ "-y", "mcp-remote", "https://mcp.host-tracker.com/mcp", "--header", "Authorization:${HT_AUTH}" ], "env": { "HT_AUTH": "Bearer YOUR_HOSTTRACKER_API_TOKEN" } } } }

The ${HT_AUTH} indirection is deliberate: some mcp-remote builds split an argument on its first space, which breaks a literal Authorization: Bearer .... Node.js 18 or newer is required.

Cursor

~/.cursor/mcp.json for every project, or .cursor/mcp.json for one:

{

"mcpServers": { "hosttracker": { "url": "https://mcp.host-tracker.com/mcp", "headers": { "Authorization": "Bearer YOUR_HOSTTRACKER_API_TOKEN" } } } }

VS Code (GitHub Copilot agent mode)

.vscode/mcp.json in the workspace. The inputs block keeps the token out of the file, prompting for it once and storing it in the editor's secret storage:

{

"inputs": [ { "type": "promptString", "id": "ht-token", "description": "HostTracker API token", "password": true } ], "servers": { "hosttracker": { "type": "http", "url": "https://mcp.host-tracker.com/mcp", "headers": { "Authorization": "Bearer ${input:ht-token}" } } } }

Windsurf

~/.codeium/windsurf/mcp_config.json:

{

"mcpServers": { "hosttracker": { "serverUrl": "https://mcp.host-tracker.com/mcp", "headers": { "Authorization": "Bearer YOUR_HOSTTRACKER_API_TOKEN" } } } }

Run it yourself (Docker)

The hosted endpoint is the normal way to use the server. If you would rather run your own copy - to read the code, audit it, or keep the MCP hop inside your network - the repository builds it from source:

docker build -t hosttracker-mcp https://github.com/HostTracker/mcp.git

docker run --rm -p 8080:8080 hosttracker-mcp

Your copy then answers at http://localhost:8080/mcp and takes exactly the same Authorization: Bearer header: it is a stateless bridge, so it stores nothing and still talks to the public HostTracker API v2 under your token. Without Docker, dotnet run --project src does the same on any machine with the .NET 10 SDK.

The same binary also speaks stdio, for clients that launch the server as a child process instead of connecting to a URL. Add --stdio and pass your token as the HT_TOKEN environment variable (there is no request header on stdio):

HT_TOKEN=YOUR_HOSTTRACKER_API_TOKEN dotnet run --project src -- --stdio

docker run -i --rm -e HT_TOKEN=YOUR_HOSTTRACKER_API_TOKEN hosttracker-mcp --stdio

Logs go to stderr in that mode, so stdout stays a clean protocol stream.

ChatGPT and other clients

ChatGPT (developer mode -> connectors): add https://mcp.host-tracker.com/mcp; ChatGPT offers its "link account" step, which opens HostTracker's sign-in + consent page. Any other client with an OAuth-capable connector dialog works the same way - just the URL.

Any client that can send a static header also works with a bearer token: the endpoint https://mcp.host-tracker.com/mcp and the header Authorization: Bearer YOUR_HOSTTRACKER_API_TOKEN. Where a client's connector form offers an API-key or custom-header authentication mode, the token goes there.

A generic, dependency-free bridge for anything that can only launch a command:

npx -y mcp-remote https://mcp.host-tracker.com/mcp --header "Authorization:${HT_AUTH}"

Longer walkthroughs, verification commands and troubleshooting live in CLIENT.md.

What the assistant can do

The server exposes the HostTracker v2 REST API as MCP tools. Every list tool takes limit (maximum 50) and cursor and returns the next cursor; every timestamp is Unix seconds in both directions; ids are opaque strings.

The full tool list

Three behaviours worth knowing before the first call:

• Bulk operations validate first. A bulk tool returns a validation report; the write needs an explicit second call with submit=true. Bulk deletion additionally needs confirmed=true and the count the validation pass reported, so a selection that drifted in between is refused.
• Bulk operations and reports are asynchronous. They answer with a job id; poll it with wait_for_job.
• Deleting anything is not undoable, so it always takes two calls. A delete tool's first call removes nothing - it returns the resource so the assistant can show you what is about to go - and only a repeat call with confirmed=true deletes. The API returns a receipt listing what was removed.

Authentication and scopes

Two ways in, same permission model:

• OAuth (connected apps). The client registers itself, you sign in once and approve a scope set on the consent page, and the server mints short-lived access tokens (1 hour) with rotating refresh tokens (90 days) behind the scenes. If the client requests no scopes it gets check + monitor:read. The account family is never grantable to a connected app. Every connection is listed under Integrations -> API -> Connected apps, where one click revokes it (the app then has to ask you again). A tool that needs a scope th