Also available as an agent skill on Agensi. The skill teaches your agent the full reforge workflow (scan → score → plan → fix → commit). Install reforge-mcp first, then add the skill.

Available on Glama - https://glama.ai/mcp/servers/Codewithaiyan/reforge-mcp

reforge-mcp

An MCP server that gives Claude Code structured tools over your codebase — dead code, duplicates, god files, the dependency graph — and tells you what it could not read.

Reports expose their limits: the health result's coverage is the fraction of counted source files analysed, summary.parse_failures counts files the parser could not read, and dep_graph.unresolved_relative_imports counts unresolved relative imports of supported source files. get_health_score returns score: null when no files were analysed. Coverage is not test coverage, and unsupported source extensions are counted from a fixed list.

The skill source is in skills/reforge/SKILL.md.

What it does

Repos built quickly with AI can accumulate unused functions, copied logic, oversized modules and circular imports. reforge-mcp gives Claude Code structured findings so it can inspect the source, prioritize changes, apply patches with optional tests, and create local Git commits. Findings and risk scores are heuristics that need review before deleting or changing code.

The 10 MCP tools

These are the exact registered MCP names; Python wrapper functions have a _tool suffix.

State lives in reforge-state.json. State writes rotate up to five timestamped backups. infer_architecture is a writing tool despite doing analysis first. The server entry point also adds Reforge entries to .gitignore and removes top-level *.tmp files in its working directory at startup.

Status

Verified against 21 open-source repositories across three validation sweeps — django, flask, fastapi, scrapy, pydantic, requests, express, axios, react, vue, svelte, zod, nest, cobra, gin, caddy, prometheus, okhttp, ktor, nowinandroid and ripgrep. Those sweeps found and closed 17 defects, including two languages whose tree-sitter queries had never compiled at all and a Python dependency graph that had been mostly missing for nine releases. Every one is reproduced and documented in IMPROVEMENTS.md.

CI runs lint and the test suite on Linux and Windows with Python 3.11, 3.12 and 3.14.

What is still true, and worth knowing before you trust a number:

• A Go dependency graph is roughly 1% resolved. Go imports are module paths that need go.mod to resolve; caddy reports 31 edges from 3,321 imports. Treat Go dependency and risk results as incomplete.

• A whole-repo scan of a large project is summarised, not complete. Long lists are capped to reduce the context needed — a 3,000-file repository comes back around 170 KB instead of 1.2 MB. Every count stays exact, and truncated names each list that was shortened with its real length, so a short list is never mistaken for a complete one.

• Dead-code detection is name-based, with no type resolution. It is deliberately biased toward missing a finding rather than inventing one, because the advice is "delete this".

• Successful patches preserve recovery information. write_fix rejects mismatched diff hunks and keeps a byte-exact .reforge-bak copy after a successful patch when the file was modified, untracked or outside Git. When Git already holds the original contents, backup is null.

Also included

• Health score trending — every get_health_score call appends to health_history so you can chart improvement over time

• Persistent memory — reforge-state.json survives across Claude Code sessions; stores the architecture hypothesis, fix log and pending work

• Monorepo support — scan_repo finds subprojects from pyproject.toml, setup.py, package.json, tsconfig.json and go.mod. It suggests a conventional test command; it does not verify that command or detect Kotlin/Gradle subprojects. When markers overlap, the first matching marker wins.

• Changelog generation — REFORGE_CHANGES.md is written on demand with score deltas, applied fixes, and pending items

Related MCP server: codex-dobby-mcp

Installation

Prerequisites: Python 3.11+, Git, and Claude Code or Claude Desktop.

API keys: Reforge does not require ANTHROPIC_API_KEY or another model API key. Its analysis and state tools run locally without calling an external model; Claude Code or Claude Desktop manages its own model access. Test commands launched by write_fix and Git hooks invoked by git_commit may have their own network activity.

Claude Code

pip install reforge-mcp
claude mcp add reforge-mcp -- reforge-mcp

Run claude mcp list to check that reforge-mcp is registered, then start a new session to use its ten tools. The executable must be on the client's PATH. To request an upgrade, run pip install --upgrade reforge-mcp; the selected version also depends on your Python version and configured package index.

Claude Desktop

pip install reforge-mcp

Add this to claude_desktop_config.json, merging it with any existing mcpServers entries:

{
  "mcpServers": {
    "reforge-mcp": {
      "command": "reforge-mcp",
      "args": []
    }
  }
}

The config is in ~/Library/Application Support/Claude/ on macOS or %APPDATA%\Claude\ on Windows. Restart Claude Desktop after saving. Its process must be able to find reforge-mcp on PATH; a terminal's activated virtualenv does not automatically change a desktop application's environment.

From source

git clone https://github.com/Codewithaiyan/reforge-mcp
cd reforge-mcp
python3 -m venv venv
source venv/bin/activate          # PowerShell: .\venv\Scripts\Activate.ps1
pip install -e .

For a source install in a virtualenv, the executable is venv/bin/reforge-mcp (Windows: venv\Scripts\reforge-mcp.exe). If the client cannot find it, use its absolute path as the Claude Desktop command or after -- in the Claude Code registration command. which reforge-mcp (PowerShell: Get-Command reforge-mcp) shows the executable from an activated environment.

Usage

Example session with Claude Code

Illustrative session; the counts and commit hash below are examples.

You: Use reforge-mcp to scan my repo at /home/me/myproject and fix the top issues.

Claude: I'll start with a full scan.

[calls scan_repo("/home/me/myproject")]

The scan found:
- 14 dead functions across 6 files
- 3 duplicate implementations of `calculate_discount`
- 1 god file: src/api/handlers.py (847 lines, 23 exports)
- 1 circular dependency: auth → middleware → auth

[calls get_health_score(repo_path="/home/me/myproject")]
Health score: 47/100

Let me fetch the duplicate implementations to confirm they're identical.

[calls get_chunk(file_path="/home/me/myproject/src/billing/utils.py", repo_root="/home/me/myproject", start_line=12, end_line=28)]
[calls get_chunk(file_path="/home/me/myproject/src/cart/helpers.py", repo_root="/home/me/myproject", start_line=5, end_line=21)]

Confirmed — identical bodies. I'll keep the one in src/billing/utils.py and have
src/cart/helpers.py import it.

[calls write_fix(file_path="/home/me/myproject/src/cart/helpers.py", repo_root="/home/me/myproject", diff_str="...", test_command="pytest")]
[calls git_commit("/home/me/myproject", files=["src/cart/helpers.py"], message="refactor: drop duplicate calculate_discount")]

Committed: a3f91c2. Running next fix...

Sample scan_repo output

Illustrative excerpt from a four-file repository; additional summary and graph diagnostics are omitted here for brevity. The example has calculate_discount copy-pasted into src/pricing.py and src/cart.py, an unused parse_legacy_csv and _old_validate, and auth.py and middleware.py importing each other.

{
  "summary": {
    "total_files": 4,
    "total_functions": 7,
    "total_classes": 0,
    "total_imports": 3,
    "total_lines": 41,
    "languages_detected": ["python"]
  },
  "dead_code": [
    { "symbol": "_old_validate",    "file": "auth.py",        "line": 8 },
    { "symbol": "checkout",         "file": "src/cart.py",    "line": 10 },
    { "symbol": "parse_legacy_csv", "file": "src/pricing.py", "line": 7 }
  ],
  "duplicates": [
    {
      "hash": "0a764010605b675494ce6d70c89bb5da7bf02f63589dd69d232c1b057c76569f",
      "locations": [
        { "file": "src/cart.py",    "line": 4, "name": "calculate_discount" },
        { "file": "src/pricing.py", "line": 1, "name": "calculate_discount" }
      ]
    }
  ],
  "god_files": [],
  "dep_graph": {
    "nodes": ["auth.py", "middleware.py", "src/cart.py", "src/pricing.py"],
    "edges": [
      { "from": "auth.py", "to": "middleware.py" },
      { "from": "middleware.py", "to": "auth.py" }
    ],
    "circular": [["auth.py", "middleware.py", "auth.py"]]
  },
  "subprojects": []
}

checkout is a dead-code candidate because nothing in those four files references its name. A finding is a question for Claude to check, not a deletion instruction. On a monorepo, subprojects fills in from the marker files it finds:

{
  "subprojects": [
    { "root": "frontend",        "language": "javascript", "test_command": "npm test" },
    { "root": "services/worker", "language": "go",         "test_command": "go test ./..." }
  ]
}

Paths use your platform's separator, so they come back with backslashes on Windows.

Sample REFORGE_CHANGES.md

# Reforge Changes

*Generated: 2025-05-02T14:33:00+00:00*

**Health score:** 73.5 (+26.5 from previous scan)

## Fixes Applied

- refactor: consolidate duplicate calculate_discount
- fix: remove dead parse_legacy_csv from src/importers/csv.py
- refactor: split handlers.py into auth_handlers.py and order_handlers.py

## Pending Fixes

- resolve circular dep: src/auth/__init__.py ↔ src/middleware/auth.py
- remove _old_validate from src/models/user.py

How it works

Reforge works in three phases:

1. Scan

scan_repo walks the repository with parse_directory, feeding each source file to the appropriate tree-sitter adapter (Python, JavaScript/TypeScript, Go or Kotlin). Tree-sitter builds a full AST — unlike regex, it understands nested scope, decorators, arrow functions, and generics. The adapters extract functions, classes, and imports into typed dataclasses.

When Git reports a nonempty tracked-file list, only those files are analysed; outside Git or with an empty index, all eligible files are analysed. languages can filter supported languages. include_tests is currently accepted but does not filter test files, whether true or false.

A missing root or a root that is a file returns an explicit error report. Existing directories without Git are still supported, including empty directories and extracted source archives.

Each parsed file also collects reference names from the AST while skipping each definition's own name and the contents of comments and strings. From the parse results, three analyses run:

• Dead code — functions whose name appears in nobody's reference set. Matching is by name, not resolved symbol, so a repo-wide name collision hides a finding: a missed one costs nothing, a wrong one costs you working code. Exempt: dunder methods, conventional entry points, test_*, anything listed in __all__, and anything carrying a decorator that is not a pure modifier like @property — a framework calls its registered functions without naming them anywhere a parser can see

• Duplicates — functions whose normalised bodies share the same SHA-256 hash. One-line methods that several classes implement identically are treated as interface conformance, not copy-paste

• God files — non-test files over 500 lines or with more than 10 top-level definitions. Methods do not count towards that total

A dependency graph is built from resolved imports — from x import y, bare import x, relative from ..pkg import y, and JS/TS ./path forms — and checked for cycles with DFS. Imports that resolve outside the repository are external and get no edge. Monorepo subprojects are discovered by scanning subdirectories for marker files (pyproject.toml, package.json, go.mod, tsconfig.json).

2. Plan

Claude Code receives the structured scan report and uses get_chunk to read specific line ranges before deciding what to fix and in what order. infer_architecture saves its hypothesis, and get_health_score appends history. The agent must use write_memory to maintain its fix log and pending queue; patches and commits do not update those lists automatically. read_memory retrieves saved values in later sessions.

3. Fix

write_fix validates a unified diff before writing the target:

1. Creates a unique byte-for-byte backup ending in .reforge-bak beside the existing target file

2. Parses the diff into hunks and checks every hunk's context and removed lines against the file at the position its header claims

3. Applies the hunks with line-offset tracking — or, if any hunk failed that check, writes nothing at all

4. Runs the optional test command (validated against the allowlist)

5. On a mismatched diff or test failure, attempts to restore the backup and returns rolled_back: true. This flag records the rollback path, not independent verification that restoration succeeded; filesystem failures can prevent recovery.

git_commit stages the listed files and commits the Git index via GitPython. It returns an error when the staged index has no changes to commit; a refused empty commit does not consume the session budget. Previously staged changes are also included, so review the index first; files_committed echoes the requested file list, not the full commit contents. The tool reads session_budget and confirm_every from reforge.toml: commits are blocked when the budget is reached, and the next call after each confirm_every group returns needs_confirmation: true until retried with confirmed: true. The count is held per repository for the life of the server process, so restarting the server starts a new session. This tool does not push.

Security

Contributing

# Clone and install
git clone https://github.com/Codewithaiyan/reforge-mcp
cd reforge-mcp
python3 -m venv venv && source venv/bin/activate
pip install -e ".[dev]"

# Run the full test suite
pytest

# Run with coverage
pytest --cov=src --cov-report=term-missing

Project structure

reforge-mcp/
├── src/reforge_mcp/
│   ├── server.py              # FastMCP server — registers all 10 tools
│   ├── tool_results.py        # TypedDict output schemas for MCP results
│   ├── tools/
│   │   ├── scan.py            # scan_repo, get_health_score, monorepo detection
│   │   ├── chunk.py           # get_chunk — line-range file reads
│   │   ├── fix.py             # write_fix — diff apply + rollback
│   │   └── git.py             # git_commit — atomic commits with budget enforcement
│   ├── scanner/
│   │   ├── parser.py          # tree-sitter adapters (Python, JS/TS, Go, Kotlin) + reference collection
│   │   ├── dead_code.py       # unreferenced symbol detection
│   │   ├── duplicates.py      # body-hash duplicate detection
│   │   ├── architecture.py    # project-shape inference
│   │   ├── risk.py            # per-symbol change risk
│   │   └── adapters/          # per-language project-detection stubs
│   └── utils/
│       ├── state.py           # reforge-state.json load/save, gitignore, changelog
│       ├── security.py        # path validation, file guard, command allowlist
│       └── diff.py            # unified diff parser and applier
├── tests/                     # one file per module, plus test_end_to_end.py
│   ├── test_parser.py         #   which drives scan -> fix -> commit for real
│   ├── test_dead_code.py
│   ├── test_duplicates.py
│   ├── test_dep_graph.py
│   ├── test_god_files.py
│   ├── test_diff.py
│   ├── test_fix.py
│   ├── test_git.py
│   ├── test_chunk.py
│   ├── test_memory.py
│   ├── test_monorepo.py
│   ├── test_scanner.py
│   └── test_security.py
├── pyproject.toml
└── reforge.toml               # optional per-repo config

reforge.toml reference

[scan]
ignore = ["generated", "vendor"]   # added to the built-in skip list

[fix]
confirm_every  = 5       # pause for confirmation every N commits
session_budget = 20      # max commits per session, per repo, before a hard stop

Those three keys are the whole file. Every one of them is read.

Changelog

See CHANGELOG.md. If you used any earlier version, re-scan. Three validation sweeps changed what the analyses report: TypeScript and Go parsed nothing at all before 0.3.0, Python dependency graphs were mostly missing before 0.4.0, and dead-code findings in Go, Kotlin and TypeScript included whole classes of false positive.

License

MIT — see LICENSE.