Inspiration
AI coding assistants sometimes suggest packages that were never published. Researchers call the attack that follows slopsquatting: someone registers the made-up name and waits for the install. Typosquatting works the same way with names one letter away from popular packages. These names end up in Jira first, in setup plans and task descriptions that an assistant drafted, long before anyone reviews a lockfile. I study software supply-chain security, and I found nothing in the Atlassian Marketplace that checks whether a package named in a work item exists at all.
What it does
- Issue panel. Open Package Reality Check on a work item. It finds every npm and PyPI package named in the summary, description and comments (install commands for npm, pnpm, yarn, pip, poetry and uv, requirement lines, package.json blocks, JS imports), asks the registries about each one and shows a verdict per package: missing, suspicious or ok, with the reasons.
- Dependency Doctor in Rovo. Ask the agent to check a work item and it explains each problem and names what was probably meant (jsonwebtoken, not jsonwebtokn; requests, not reqeusts). It never calls a package safe, only says what the evidence shows.
- MCP tools for coding agents. The best moment to catch a made-up package is before the agent installs it. Through the Atlassian MCP server (rovo:mcp), any MCP client a site admin allows can call the same two checks with the user's own Jira permissions. For agents that cannot do the OAuth sign-in, the Jira admin page also shows a private MCP URL served by the app itself, with a one-line command for Claude Code, Cursor or Rovo Dev; the agent then calls check_packages or scan_text on its own plan.
- Teamwork Graph. Once a site admin adds the connector, every scan writes a security remote link to the site's Teamwork Graph: the work item's ARI, the counts and each package's verdict, with a red, amber or green status.
How it decides
A package that does not exist on npm or PyPI is flagged as missing at once, and a lookalike of a popular name is suggested. For a package that exists, points add up: name one or two edits from one of the 3,000 most-used packages (3), first published under 30 days ago (2) or under six months ago (1), under 100 downloads last week (1), install scripts (1), no source repository (1), deprecated (1), known vulnerabilities on the latest release (1). Three points or more is suspicious. Packages on the popular lists are never flagged.
How I built it
Forge (Node runtime) with TypeScript. One engine (name normalisation, edit distance, registry checks, text extraction) is shared by the issue panel resolver, the two Rovo actions (which the rovo:mcp module also exposes as MCP tools), the MCP web trigger and the Teamwork Graph connector. The popular-name lists come from npm-high-impact (top 2,000) and hugovk/top-pypi-packages (top 1,000). The web trigger endpoint is stateless JSON-RPC over HTTP POST: initialize, ping, tools/list and tools/call, with batches. Egress is limited to registry.npmjs.org, api.npmjs.org and pypi.org. 24 offline tests plus a live smoke test against the real registries.
Challenges
- Teamwork Graph has no vulnerability object type, so each scan is a remote link of type security. The documented examples did not match what the ingestion API accepts: it wanted schema version 2.0, a single permissions object and string-only attributes, and it rejects any association from a remote link to a Jira work item. I found the accepted shape by sending variants to the live API, and the link now names its work item by ARI and URL.
- Two ways into MCP: rovo:mcp gives admins control and per-user permissions, while the app's own endpoint works for agents that cannot complete an OAuth sign-in. Both call the same engine.
What I learned
Most of the risk sits in plain text: a package name in a task description is trusted the moment someone copies the command. Checking existence and name distance catches the made-up and lookalike names without any secret, account or code host.
What's next
More ecosystems (crates.io, RubyGems, Go modules), a Bitbucket pull request check, and a per-site allow list for internal package names.
Privacy
Only package names leave Atlassian, to npm and PyPI. The app keeps one value in Forge storage, the Teamwork Graph connection id. Scan results are not stored by the app; when the connector is on, they are written to the site's own Teamwork Graph.
How it meets the Forge requirement
Built entirely on Forge: jira:issuePanel, jira:adminPage, rovo:agent and rovo:action, rovo:mcp, webtrigger (MCP), graph:connector, a Jira issue trigger, and Forge storage. No external backend.
Demo
Narration in the video is text-to-speech. Every screen is the real app on a demo site.
Built With
- forge
- jira
- mcp
- node.js
- npm
- pypi
- rovo
- teamwork-graph
- typescript
- vitest