Claude Code quickstart: check what changed in your code

Install the plugin or the server

We suggest the plugin. It adds the MCP server plus four skills that tell Claude when to check. Run these in a Claude Code session:

/plugin marketplace add since-dev/mcp
/plugin install since-dev@since-dev

The install opens the plugin's details, where you pick who gets it: you, everyone on this repository, or you in this repository only (Claude Code plugin docs). If the summary asks you to, run /reload-plugins. For the server alone, run this in your shell:

claude mcp add --transport http since \
  https://api.since.dev/mcp

Add --scope user for every project, or --scope project to share it through .mcp.json (Claude Code MCP docs).

Sign in with OAuth

1. Run /mcp in Claude Code and pick the Since.dev server.
2. In your browser, sign in to Since.dev (a free account works) and approve Claude Code for one workspace.
3. Back in Claude Code, /mcp shows the server connected, with three tools: whats_changed, check_code and fix_code.

No key goes in a file, and Claude Code refreshes the sign-in for you. To use another workspace, choose Clear authentication in the /mcp menu and sign in again (MCP reference).

Three things to ask first

Ask in plain words. Claude picks the tool, and every answer cites its source.

Which AI models in this repo are retiring?

Claude sends the files that call a model to check_code. Each model comes back with its status, its date on the platform your code calls, the days left and the replacement. A shortened example:

src/llm.ts line 1: claude-sonnet-4-5-20250929
  status: deprecated
  retires: 2026-11-30 on the Claude API
  replacement: claude-sonnet-5-5

What breaks if I upgrade this package?

Claude checks your lockfile, then asks whats_changed with since set to your version. You get each breaking change, deprecation and advisory since then, with what to use instead. Have it check the files that use the package too, and you get the lines. An example, for a made-up package:

acme-sdk: 3.2.0 installed, 4.1.0 latest
removed in 4.0.0: Client.send
  use instead: Client.request
  source: the 4.0.0 release notes

What changed since you were trained?

Before Claude writes code against a library it learned in training, have it call whats_changed with since set to a date or cutoff:<model>. You get renamed and removed functions, changed parameters and retired models, each linked to the publisher's page.

Treat needs_review and unknown as not settled, never as safe. Read the cited source or decide yourself.

What the plugin's skills add

The skills tell Claude when to check, so you don't have to ask. Claude loads one when your task matches its description, or you can run one, such as /since-dev:since-before-upgrading (skills docs).

Fixes, with your approval

fix_code writes, so Claude Code asks before it runs. Give it a snippet and it returns the patch, on Team and up. Give it a connected repository and it opens fix pull requests, on Business and Platform. Fixes count against the month's fix pull requests, not credits.

To skip approving the two read tools, allow mcp__since__whats_changed and mcp__since__check_code with /permissions. The plugin's names start with mcp__plugin_since-dev_since__ (permission rules).

Use an API key for CI and headless runs

OAuth needs a browser. For CI or claude -p where nobody can sign in, use a workspace API key:

1. In app.since.dev, pick the workspace, open API keys and choose Create API key. It starts with sc_live_ and is shown once.
2. Store it as a CI secret, such as SINCE_DEV_KEY, and add the server with it as a header:

claude mcp add --transport http since \
  https://api.since.dev/mcp \
  --header "Authorization: Bearer $SINCE_DEV_KEY"

A key reaches only its own workspace. Never commit one, and revoke it under API keys when you're done. Without an agent, the HTTP API takes the same arguments.

What calls cost

Calls spend credits from your workspace's monthly allowance. Every plan includes some, Free too, and every response says how many are left.

fix_code costs what its check costs. A call that fails costs nothing, and so do a check's later pages read with nextCursor within 15 minutes (credits). Prices change, see since.dev/#pricing.

Troubleshooting

• claude mcp list shows each server's health. For Failed to connect, claude mcp get since gives the reason.
• Needs authentication: run /mcp and sign in. If no browser opens, copy the URL Claude Code shows.
• No skills after the plugin install: run /reload-plugins.
• A 401 means the sign-in expired or the key isn't valid. Sign in again or make a new key.
• credits_exhausted means the month's credits are spent. They reset monthly.

Where Since.dev fits

Since.dev tracks breaking releases, deprecations, model retirements and end-of-life dates, and matches each one to the file and line it hits.

• Scan any public GitHub repository, no sign-in needed.
• See every Claude model date on the Claude retirement calendar.

Connect GitHub and, on Team and up, it opens fix pull requests for people to review. Nothing merges or deploys.

Questions

Do I need a paid plan to use Since.dev in Claude Code?

No. whats_changed and check_code work on every plan, Free included. Fixes start on Team.

Should I install the plugin or only the server?

The plugin. It adds the same server plus four skills, so Claude checks before upgrades, model changes and CI edits without being asked.

What does Claude Code send to Since.dev?

Only what the tool call holds, such as a model ID, a lockfile or a snippet. Code is read to analyze it, then deleted.

Can I use it in CI?

Yes. Add the server with a workspace API key as a bearer header, or call the HTTP API directly. Keep the key in CI secrets.

Related guides