# Plagiarism Analysis: darshil3011/codedrift vs jgravelle/jcodemunch-mcp 

## Timeline

|              | jcodemunch-mcp  | codedrift          |
| ------------ | --------------- | ------------------ |
| First commit | 2026-02-08      | 2026-04-12 (today) |
| Commit count | 500+            | 10                 |
| Version      | 1.35.1          | 0.1.0              |
| PyPI         | Yes (since Feb) | No                 |

codedrift appeared **63 days after** jcodemunch-mcp's initial commit and the same day as jcodemunch v1.35.1. The entire repo was created in a single session:

> "added files" → "added source code" → README polish — all within ~14 hours.

---

## Verdict: Concept clone, not line-for-line copy

This isn't a git clone with renamed files. It's a **clean-room reimplementation** of jcodemunch-mcp's architecture, almost certainly generated by prompting an LLM with jcodemunch's README/docs/tool descriptions and asking it to build a competing tool.

---

## Structural Evidence

### Identical architecture decisions

| Design choice                                      | jcodemunch-mcp              | codedrift             |
| -------------------------------------------------- | --------------------------- | --------------------- |
| SQLite + FTS5 for symbol search                    | Yes                         | Yes                   |
| tree-sitter for AST parsing                        | Yes                         | Yes                   |
| BM25 ranking from FTS5                             | Yes                         | Yes                   |
| Language adapter pattern per language              | Yes (LanguageSpec)          | Yes (LanguageAdapter) |
| Same 4 languages (Python, JS/TS, Go, Rust)         | Yes (+ 15 more)             | Yes (only these 4)    |
| Symbol/CallSite/ImportRef data model               | Yes (Symbol dataclass)      | Yes (identical names) |
| search + resolve + overview tool pattern           | Yes                         | Yes                   |
| Token savings tracking per call                    | Yes (token_tracker.py)      | Yes (savings.py)      |
| Session-aware diff reads                           | Yes (session state)         | Yes (differ.py)       |
| install-skill → writes CLAUDE.md rules             | Yes (cli/init.py)           | Yes (skill.py)        |
| install-hook → git post-commit                     | Yes (hooks system)          | Yes                   |
| sentence-transformers embedding for session memory | Yes (local_encoder.py ONNX) | Yes (memory.py)       |
| all-MiniLM-L6-v2 as embedding model                | Yes                         | Yes                   |

---

## Matching tool names and descriptions

codedrift's MCP tools are renamed versions of jcodemunch's core tools:

| jcodemunch tool     | codedrift tool     | Same purpose?                               |
| ------------------- | ------------------ | ------------------------------------------- |
| search_symbols      | codedrift_search   | Yes — FTS5 symbol search                    |
| get_symbol_source   | codedrift_resolve  | Yes — full symbol context + callers + tests |
| get_repo_outline    | codedrift_overview | Yes — structural map                        |
| get_file_content    | codedrift_read     | Yes — file reads                            |
| get_session_context | codedrift_memory   | Yes — session recall via embeddings         |

---

## Matching marketing language

codedrift's README echoes jcodemunch's value proposition nearly verbatim:

* "Eliminate token waste in AI coding agents"
* "90% of tokens finding code, not writing it"
* "50x reduction claim"
* "No LLM involved in indexing — tree-sitter is a deterministic AST parser"
* Tool priority rules:

  * "Use BEFORE Grep or Glob"
  * "Use INSTEAD of reading full files"

---

## What codedrift lacks (proving derivation)

codedrift implements only the **v0.x core** of jcodemunch — the simplest subset:

* Missing 40+ tools:

  * No blast radius, call hierarchy, dead code detection, dependency graph, coupling metrics, hotspots, refactoring planner, class hierarchy, churn rate, layer violations, extraction candidates, etc.
* Missing 15+ languages:

  * No Java, PHP, C/C++, Swift, Dart, C#, Ruby, Elixir, Erlang, Fortran, SQL, Kotlin, Scala, etc.
* No async indexing (jcodemunch uses `asyncio.to_thread`)
* No GitHub remote indexing (jcodemunch has `index_repo` via httpx)
* No framework context (Django, Express, Rails, Laravel, Next.js, Nuxt, Go routers)
* No LSP bridge for compiler-grade resolution
* No cross-repo support
* No config system, no security/path validation, no rate limiting
* No AI summarization tier
* No PageRank/signal fusion for search ranking
* No complexity metrics (cyclomatic, nesting, param count)
* No watcher system for live reindexing
* No hooks (PreToolUse, PostToolUse, PreCompact, etc.)
* No tests to speak of (3 test files vs 2995 tests)
* No CHANGELOG, no versioning discipline

> This pattern — exact same architecture, exact same core features, none of the depth — is the signature of someone who studied the public API surface and README but never read the actual implementation.

---

## Author Profile

* darshil3011 (Darshil Modi) — GitHub account since ~2019 (based on user ID)
* No other MCP-related repos
* 6 stars on day one (suspicious — likely self-promotion or coordinated)
* No attribution to jcodemunch anywhere in the repo

---

## Risk Assessment

| Factor                     | Level                                  |
| -------------------------- | -------------------------------------- |
| Line-for-line code copying | Low — rewritten, not copied            |
| Architectural copying      | High — identical design at every layer |
| Marketing copying          | High — same taglines, same claims      |
| Commercial threat          | Low — no PyPI, no tests, no depth      |
| Attribution                | None given                             |

---


## Bottom Line

Someone used an LLM to clone jcodemunch's architecture in a day.

It’s flattering.
It’s ethically questionable.
It’s not legally actionable.

And with a **63-day head start, 2995 tests, and 40+ tools**, it’s not a competitive threat.