MCP server for GitHub issues
The GitHub issues worth your day, with every point itemized.
Ask your AI assistant what to work on today. IssueWatcher ranks fresh issues from the repos you starred and the libraries your projects depend on, then shows how each score adds up.
- Read-only
- One GitHub token, no AI key
- MIT licensed
- v0.2.5
Scoring
Every score is a short sum.
The ranking is plain rules you can read. Each answer carries the lines that applied, so you can disagree with one line and still use the rest. Your assistant does the thinking with its own model. IssueWatcher never calls one.
| You depend on it at runtime | +40 |
| A repo you starred | +30 |
| Mentions the version you use | +25 |
| Title says crash, regression or CVE | +20 |
| Labelled good first issue | +20 |
| Opened in the last 48 hours | +10 |
| Unassigned | +10 |
| Has reproduction steps | +5 |
| Already assigned | −25 |
| Vague title, like “Report a bug” | −20 |
| A team is already on it | −20 |
| A fix is already planned or open | −15 |
| Maintainer chore: CI, releases, bots | −15 |
| Title is about a newer major than yours | −15 |
| Open 14 days, no reply, no taker | −10 |
| Feature request nobody has answered | −10 |
Values for the default mode, as of version 0.2.5. Some lines apply only in one mode, and the total is held between 0 and 100.
contributor
What could I pick up and fix? Unassigned, fresh and not already claimed.
consumer
What could break my projects? Issues in libraries you use, matched to your versions.
both
A mix of the two. This is the default.
Tools
Three tools. All read-only.
Ask in plain words. Your assistant picks the tool and reads the result to you.
get_ranked_issuesWhat GitHub issues should I work on today?
Ranks recent issues across your starred repos and your dependencies. Choose the mode, a window of up to 30 days, a minimum score, or only certain repos.
explain_issueWhy did you rank the first one so high?
Lists every signal with its points and how the issue relates to your repos. A short excerpt of the issue text comes only when you ask, and is labelled as untrusted.
list_watched_reposWhat repos are you watching for me?
Shows the repos you starred and the source repos of the libraries your projects use, with which of your projects use which.
Install
Install in about two minutes.
You need one GitHub token with read-only access. No AI key and no account.
- Create a token. Open the fine-grained token page, name it IssueWatcher, and choose public repositories, read-only. To include dependencies from your private repos, select them and set Contents to read-only.
- Add IssueWatcher to your AI app, using the tabs.
- Ask a question from the box below the tabs.
Treat the token like a password. Never paste it into a chat.
- Download issuewatcher-0.2.5.mcpb from the latest release.
- Open the file. Claude Desktop asks for your token and keeps it in the system keychain.
- Turn the extension on and start a new chat.
One click, and nothing else to install.
Run this once, then start claude in the same folder.
claude mcp add issuewatcher -e GITHUB_TOKEN=github_pat_xxxxxxxx -- npx -y issuewatcher-mcp
In Windows PowerShell, put the double dash in quotes: '--'. For a start that takes under a second, run npm install -g issuewatcher-mcp once and use issuewatcher-mcp in place of npx -y issuewatcher-mcp.
Add this to ~/.cursor/mcp.json.
{
"mcpServers": {
"issuewatcher": {
"command": "npx",
"args": ["-y", "issuewatcher-mcp"],
"env": { "GITHUB_TOKEN": "github_pat_xxxxxxxx" }
}
}
}
VS Code, Windsurf, Zed and others work the same way. See the setup guide.
Then ask
What GitHub issues should I work on today?
If your app answers with general advice, add “use IssueWatcher” to the question.
Safety
What it can and cannot touch.
Read-only
No tool can comment on, assign, label or change anything on GitHub. The token you give it can be read-only too.
Your token stays on your computer
It goes to GitHub and nowhere else. It never appears in logs or in answers.
No telemetry
IssueWatcher sends nothing to its author. It talks to api.github.com, and to the npm and PyPI registries to find where your dependencies live.
Strangers write issues
A malicious issue could try to steer an AI. IssueWatcher strips hidden characters, link targets and markup, caps lengths, and leaves issue bodies out unless you ask for a labelled excerpt. That lowers the risk and cannot remove it.
Status
Early, and checked in small ways.
Version 0.2.5. The ranking is hand-written rules, tuned by reading real results.
First check: 99 real issues from one developer's watch list were labelled useful or noise by a model, with the scores hidden. IssueWatcher's ranking separates the two with an AUC of 0.82, where 1.0 is perfect and 0.5 is chance.
That is a first look, not proof. The band is wide because there are only 99 issues. Whether the top 10 beats newest-first is not established, and no people have labelled yet.
- Used by hand on Windows with Node 22, in Claude Desktop and Claude Code.
- The test suite also runs on Linux in CI.
- Cursor, VS Code and macOS are untested.
If a ranking looks wrong, open an issue with the issue number and what you expected.