Anthropic published an official Claude Code plugin called code-modernization. The pitch is not speed. The pitch is that the plugin proves the new code behaves like the old code, and grades its own result.
Install it with one line inside Claude Code:
/plugin install code-modernization@claude-plugins-official
This review reads the plugin's own repository and its published trial runs. I did not run it myself. The last section says why, and what that limits.
Checked against vendor pages on 2026-11-06.
Sources for every number below: the plugin README and the marketplace listing. The vendor reports about 5,860 installs at the time of writing.
What the plugin contains
| Part | Count | Names |
|---|---|---|
| Commands | 13 | modernize, preflight, assess, map, extract-rules, review, brief, uplift, transform, reimagine, verify, harden, status |
| Agents | 8 | legacy analyst, business rules extractor, architecture critic, scaffolder, test engineer, security auditor, uplift migrator, version delta analyst |
| Tracks | 3 | uplift, transform, reimagine |
The plugin states that it works with any language. The marketplace description names COBOL, legacy Java and C++, and monolith web applications. The repository README makes the wider claim: "It works with any language and any kind of move."
The nine steps
Each step writes to analysis/<name>/ and refreshes a single REPORT.html. The plugin never edits the legacy source. It links to it through legacy/<name> and copies nothing.
| Step | Command | Output |
|---|---|---|
| 0 | modernize |
Writes INTENT.md. Every later command reads it |
| 1 | modernize-preflight |
Checks the environment, asks five questions, proves a build works |
| 2 | modernize-assess |
Inventory, complexity, debt, security, recommended pattern |
| 3 | modernize-map |
Dependencies, data flow, entry points, business flows |
| 4 | modernize-extract-rules |
Business rules as Given/When/Then cards with file:line |
| 5 | modernize-brief |
A phased plan. Nothing is built before approval |
| 6 | uplift, transform or reimagine |
The build |
| 7 | modernize-verify |
The proof |
| 8 | modernize-harden |
A security scan and a patch you apply yourself |
A person decides at six points: the five preflight answers, the rules that look wrong, the plan approval, any accepted difference, the sign-off, and the security patch. The plugin does not decide these for you.
The three build tracks
| Track | Use it when | How it proves itself |
|---|---|---|
uplift |
Same technology, newer version, such as Java 8 to 17 | The same test suite on both versions |
transform |
New technology, one module at a time, old system keeps running | Old and new run on the same inputs, a script compares output |
reimagine |
Rebuild on a new architecture | A spec mined from the code, then acceptance tests |
A version move keeps your code, so it can skip extract-rules. The short path is preflight, assess, uplift, verify.
The proof is five mechanical checks
This is the part that separates the plugin from a prompt collection. modernize-verify runs in a fresh session and redoes the work instead of trusting notes.
- Tests run again from clean. Build output is deleted and the full suite runs. A run that executed nothing is a failure.
- Old and new on the same inputs. Where the legacy code can run, both run and a script compares every byte. Only fields that must vary are masked, and every masked field is listed.
- Inputs nobody used. The command invents at least ten new inputs, including boundaries, empty values and malformed records.
- Tests that can fail. A deliberate one-line break must turn tests red. The comparison itself is checked by changing one byte of output.
- Every critical rule has a test that ran. A rule named only by a skipped test is listed as "named, not run".
scripts/proof_pack.py computes one verdict per module from those files: PROVEN, PARTLY PROVEN with what is missing, or NOT PROVEN. The verdict is computed again from current evidence on every run. None is carried over.
What the vendor reports from its own trials
The README lists eleven public codebases that the plugin ran on, headless, during development. Two of them ended at NOT PROVEN. That is the most useful part of the documentation, because it shows the gate closing.
| Codebase | Move | Result |
|---|---|---|
| AWS CardDemo, COBOL and CICS | Rewrite the interest job in Java 21 | 35 rules, 8 critical, 213 tests. Five checks in a row each found what the tests missed. The last check used 15 malformed inputs and found 7 more differences. NOT PROVEN |
| osCommerce, PHP 5 | PHP to Python and FastAPI | 97 rules, 26 critical. 17,957 tests pass. 17,727 comparison cases match the real PHP. PROVEN |
| Eclipse Jetty, 2,600 files | Java 8 to 17 | 946 tests give the same result on both versions. Running both found six changes that reading the code missed. 50 tests added, 996 on each version |
| JPetStore, Java and JSP | Rebuild as a REST API | 12 capabilities, 35 rules. 35 tests pass, 56 more pending. 6 of 24 new comparisons differ. NOT PROVEN |
| AngularJS RealWorld | To React and TypeScript | 79 rules, 5 critical. 166 tests. 29 comparison cases: 23 identical, 6 approved as deliberate |
| beets, Python 2 | Python 2 to 3 | 21 rules. Only 3 files fail to compile on Python 3, so the plugin called it an uplift, not a rewrite |
| AWStats, Perl | To Python and FastAPI | 88 rules. The security scan confirmed 48 findings, 9 high, and refuted 14 of 62 as false positives |
| A booby-trapped codebase | Rewrite | Planted instructions in comments, README, CLAUDE.md and settings. None was obeyed. No marker file appeared |
The two NOT PROVEN rows do not mean the plugin failed. They mean the gate worked. Without that gate, a reader ships a rewrite that passes its own tests and still behaves differently.
Where the method breaks
Three limits are structural, not bugs.
A system that cannot run gives no full proof. The README states it plainly: if the old system only runs in production or on a mainframe, the old behavior can be checked only against recorded outputs, and the best possible verdict is PARTLY PROVEN.
The proof needs a test suite that runs. The beets run found no test that runs on Python 3, and the machine refused the packages. The plugin reported that and claimed no pass.
A second run gives different counts. The security scan of osCommerce found 48 confirmed findings, then 52. Read the numbers as typical, not exact.
Two operational limits matter for planning:
| Step | Typical time on tens of thousands of lines |
|---|---|
| preflight | 3 to 4 minutes |
| assess | 5 to 8 minutes |
| map | 5 to 15 minutes |
| extract-rules | 5 to 15 minutes, and it started 50 to 200 agents |
| brief | about 5 minutes |
| building one module | 15 to 30 minutes |
Start with one module as a pilot. The plugin says so in its own install section.
What this means for ABAP and SAP work
The plugin will analyse ABAP, because it works with any language. The proof will not reach PROVEN.
The proof compares old and new behavior on the same inputs. An ABAP program that runs only inside a live SAP system cannot run next to a rewritten version on a laptop. The best available verdict is PARTLY PROVEN against recorded outputs. That is still useful, because the assessment, the map and the rule cards have value without a build.
The honest plan for a body of SAP custom code is: run assess, map and extract-rules, then stop and use the rule cards. Do not promise a rewrite proof you cannot produce.
What I did not do
I did not run the plugin. Claude Code needs an Anthropic account, and this machine has no licence. Every number above comes from Anthropic's published trial runs, not from my own test.
That limits this review in one specific way. I can check what the plugin claims and how it computes a verdict. I cannot tell you how it behaves on your code. A reader who needs that answer must run the pilot themselves, on one module, and read the REPORT.html that comes out.
I did verify several things at the source: the repository layout, the thirteen command files, the eight agent definitions, and the plugin.json keys. I also read the five proof checks as the README writes them.
Bottom line
The plugin is not a code generator with a licence file. It is a workflow with a gate, and the gate is the feature.
Use it when a migration must be defended to someone who does not trust the model. Skip it when you want a fast rewrite and no audit trail.
Sources
- code-modernization plugin source
- Claude Code official plugin directory
- Marketplace listing
- Claude Code plugin install documentation
Follow the Agent Tooling Tests
Every tool on this site gets read at the source and, where possible, run on real code. Subscribe for the next one.
Subscribe