About dep-diff-mcp
dep-diff-mcp is an MCP server for developers reviewing package upgrades in npm, PyPI, and GitHub Actions. It combines version classification, release-note changes, security fixes, migration links, and ranked recommendations.
Key features
- 01Single-package upgrade analysis
- 02Bulk analysis of up to 50 changes
- 03Semver and breaking-change detection
- 04Security fixes and migration links
- 05Risk-ranked recommendations
Use cases
- 01Reviewing a Dependabot pull request
- 02Checking risk before a version bump
- 03Prioritizing upgrades across a dependency batch
MCP support
Exposes `analyze_package_change` and `analyze_packages_bulk` to MCP clients. Responses include JSON and schema-validated structured content.
Configuration
Remote servers need no local install. Pick your client, then copy the config or install in one click.
https://dep-diff.digicatalyst.ca/mcpclaude mcp add --transport http digicatalyst-systems-dep-diff https://dep-diff.digicatalyst.ca/mcp{"mcpServers":{"digicatalyst-systems-dep-diff":{"type":"http","url":"https://dep-diff.digicatalyst.ca/mcp"}}}Open in Cursor{"servers":{"digicatalyst-systems-dep-diff":{"type":"http","url":"https://dep-diff.digicatalyst.ca/mcp"}}}Open in VS Code{"mcpServers":{"digicatalyst-systems-dep-diff":{"type":"http","url":"https://dep-diff.digicatalyst.ca/mcp"}}}{"mcpServers":{"digicatalyst-systems-dep-diff":{"serverUrl":"https://dep-diff.digicatalyst.ca/mcp"}}}Run commands in a terminal; paste JSON into the client’s MCP config and restart. Servers that need sign-in will prompt on first connect.
Tools
2| Parameter | Type | Required | Description |
|---|---|---|---|
| name | string | Yes | Package name (e.g. 'react', 'requests') |
| ecosystem | enum | Yes | Package ecosystemnpm | pypi | github-actions |
| toVersion | string | Yes | Target version (e.g. '19.0.0') |
| fromVersion | string | Yes | Current version (e.g. '18.2.0') |
Details
From the project READMEdep-diff-mcp
MCP server that translates a lockfile diff into a human-readable upgrade plan.
Point your AI assistant (Cursor, Claude Desktop, Claude Code) at a Dependabot PR, npm outdated output, or any pair of package versions, and get back a ranked upgrade plan: semver class, breaking changes pulled from GitHub release notes, CVEs fixed in the range, migration guide links, and a clear recommendation per package.
What it looks like
Ask your assistant, in plain language:
Is it safe to bump lodash from 4.17.20 to 4.17.21?
It calls analyze_package_change and gets back:
{
"package": "lodash",
"ecosystem": "npm",
"fromVersion": "4.17.20",
"toVersion": "4.17.21",
"semverClass": "patch",
"repoUrl": "https://github.com/lodash/lodash",
"releaseCount": 0,
"breakingChanges": [],
"securityFixes": [
{
"id": "GHSA-29mw-wpgm-hmr9",
"summary": "Regular Expression Denial of Service (ReDoS) in lodash",
"severity": "MODERATE"
},
{
"id": "GHSA-35jh-r3h4-6jhm",
"summary": "Command Injection in lodash",
"severity": "HIGH"
}
],
"migrationLinks": [],
"recommendation": "RECOMMENDED: 2 security fix(es) (incl. high/critical).",
"recommendationLevel": "security"
}
A patch bump you would normally merge without looking. It closes a HIGH-severity command injection. That is the case this server exists for.
GitHub Actions
Dependabot opens more actions/* pull requests than almost anything else, and its
scheduled version updates carry no advisory data. Ask:
Is it safe to bump tj-actions/changed-files from 45.0.7 to 46.0.1?
{
"package": "tj-actions/changed-files",
"ecosystem": "github-actions",
"semverClass": "major",
"repoUrl": "https://github.com/tj-actions/changed-files",
"securityFixes": [
{
"id": "GHSA-mrrh-fwg8-r2c3",
"summary": "tj-actions changed-files through 45.0.7 allows remote attackers to discover secrets",
"severity": "HIGH"
}
],
"recommendation": "RECOMMENDED: 1 security fix(es) (incl. high/critical).",
"recommendationLevel": "security"
}
A whole Dependabot batch
Here's my Dependabot PR — what's actually risky in it?
analyze_packages_bulk takes up to 50 changes at once and ranks them
security > caution > review > likely-safe > safe:
{
"totalPackages": 5,
"bySemverClass": { "major": 1, "minor": 3, "patch": 1 },
"securityFixesTotal": 7,
"packagesWithBreakingChanges": 1,
"packages": [ /* one entry per package, same shape as above */ ]
}
Condensing the recommendation field of each entry, that batch comes back in this order:
| Package | Change | Class | Verdict |
|---|---|---|---|
lodash | 4.17.20 → 4.17.21 | patch | RECOMMENDED — 2 security fixes (incl. high/critical) |
axios | 1.6.0 → 1.7.9 | minor | RECOMMENDED — 1 security fix (incl. high/critical) |
express | 4.18.2 → 5.0.0 | major | RECOMMENDED — 2 security fixes; 3 breaking changes, migration guide |
requests (PyPI) | 2.31.0 → 2.32.0 | minor | RECOMMENDED — 2 security fixes |
typescript | 5.3.3 → 5.4.5 | minor | LIKELY SAFE — minor version, additive changes per semver |
Note the ordering: the patch bump outranks the major one. Semver tells you how much
changed; it does not tell you what is urgent.
Every response above is real output from the hosted instance, trimmed only where marked.
Install
Claude Code
One command, user scope (available in every project):
claude mcp add -s user dep-diff -- npx -y @digicatalyst/dep-diff-mcp
Project scope (writes .mcp.json at repo root, team-shared):
claude mcp add -s project dep-diff -- npx -y @digicatalyst/dep-diff-mcp
With an explicit token (skip this if you have the gh CLI authenticated — see GitHub token below):
claude mcp add -s user --env GITHUB_TOKEN=ghp_xxx dep-diff -- npx -y @digicatalyst/dep-diff-mcp
Verify:
claude mcp list
Restart the Claude Code session to pick up the server.
Cursor and Claude Desktop
Add to your MCP client config:
- Cursor:
~/.cursor/mcp.json - Claude Desktop:
~/Library/Application Support/Claude/claude_desktop_config.json(macOS) or%APPDATA%\Claude\claude_desktop_config.json(Windows)
{
"mcpServers": {
"dep-diff": {
"command": "npx",
"args": ["-y", "@digicatalyst/dep-diff-mcp"]
}
}
}
Restart your MCP client. Ask something like "what's risky in this Dependabot PR?" and the tools are invoked automatically.
Hosted remote (no install)
A hosted instance runs at https://dep-diff.digicatalyst.ca/mcp over streamable HTTP, so you can skip the npm package entirely:
claude mcp add -s user -t http dep-diff https://dep-diff.digicatalyst.ca/mcp
Or in a client config:
{
"mcpServers": {
"dep-diff": {
"type": "http",
"url": "https://dep-diff.digicatalyst.ca/mcp"
}
}
}
The hosted instance authenticates to GitHub with its own token, so release-note lookups work at full rate limits without you configuring anything. Pass ?githubToken=ghp_xxx on the URL only if you would rather requests counted against your own GitHub quota. It is stateless and keeps no logs of your queries — see PRIVACY.md. Run the npm package locally instead if you would rather your token never leave your machine.
The same instance is also listed on Smithery, which proxies to it through their gateway.
GitHub token (optional but recommended)
The server hits the GitHub API to read release notes. Without a token you get 60 requests per hour (GitHub's anonymous limit) — enough for occasional single-package queries, not enough for bulk lockfile analysis.
The server resolves a token in this order:
GITHUB_TOKENenvironment variable, if set.gh auth token— if the GitHub CLI is installed and authenticated, the server uses that token automatically. No config change needed.- Anonymous (60 req/hr).
Recommended: use the gh CLI
If you already have gh installed (brew install gh && gh auth login), stop here — the server picks up your existing auth. No plaintext token anywhere.
Alternative: environment variable
Create a fine-grained token at https://github.com/settings/tokens:
- Token name:
dep-diff-mcp - Expiration: 90 days (rotate periodically)
- Repository access:
Public Repositories (read-only)— no private repo access - Permissions: none beyond the default public read — do not grant
repo,workflow,user, or any write scope
Then reference it in the MCP config:
{
"mcpServers": {
"dep-diff": {
"command": "npx",
"args": ["-y", "@digicatalyst/dep-diff-mcp"],
"env": { "GITHUB_TOKEN": "github_pat_xxx" }
}
}
}
Security notes
- This config file lives on your disk in plaintext. Keep perms tight (
chmod 600) and do not paste the token into AI chats, issues, or shared screens — transcripts are often retained. - The token in this config should be least-privilege (public repo read only). Even leaked, it can only read public data you could already read.
- Rotate tokens periodically. Revoke any token that may have been exposed at https://github.com/settings/tokens.
- The server never writes the token to stdout/stderr or the response payload.
Tools
analyze_package_change
Analyze one package upgrade. Inputs: ecosystem (npm, pypi, or github-actions), name, fromVersion, toVersion.
analyze_packages_bulk
Analyze up to 50 package upgrades in parallel. Returns packages ranked by risk (security > caution > review > likely-safe > safe), plus summary counts.
What you get back
- Semver classification — major / minor / patch / downgrade / unknown
- Breaking changes — extracted from GitHub release notes headers
- Security fixes — CVEs present at
fromVersionbut resolved attoVersion(via OSV.dev) - Migration links — upgrade guide URLs found in release notes
- Recommendation — single-line verdict + level
Both tools return this twice: as the JSON text block shown above, and as MCP
structured content
validated against a declared outputSchema. A client that supports structured output can read
recommendationLevel or securityFixes[] straight off the response instead of re-parsing the text.
Clients that don't are unaffected — the text block is unchanged.
Supported ecosystems
- npm
- PyPI
- GitHub Actions — pass the action reference as the name (
actions/checkout, or a nestedgithub/codeql-action/init). An action reference is already a repository coordinate, so release notes and breaking changes resolve without a registry lookup, and bare major tags such as4 -> 5classify correctly.
Development
npm install
npm run build
GITHUB_TOKEN=ghp_xxx npm run inspect # MCP Inspector
License
MIT
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.