CodeWalker

Walk your codebase before writing new code.

CodeWalker is an MCP server that gives Claude Code real-time access to your Python codebase structure, enabling AI-assisted development that reuses existing code instead of duplicating it.

The Problem: AI Code Duplication

What Happens Without CodeWalker

When Claude Code writes code, it can't see what already exists in your codebase. This causes a cascade of problems:

Day 1: You ask Claude to add CSV loading functionality

# Claude creates: src/data_loader.py
def load_csv_file(path):
    return pd.read_csv(path)

Day 5: Different feature needs CSV loading

# Claude creates: src/importer.py (Claude has no memory of data_loader.py)
def load_csv_data(filepath):
    df = pd.read_csv(filepath)
    return df

Day 10: Another feature, another duplicate

# Claude creates: src/utils.py (Claude still doesn't know about the others)
def read_csv(file_path):
    return pd.read_csv(file_path, low_memory=False)  # Now with different behavior!

Result after 2 weeks:

• 🔴 7 different CSV loading functions across your codebase

• 🔴 Inconsistent behavior (one uses low_memory=False, others don't)

• 🔴 Impossible to maintain (bug fixes need to be applied 7 times)

• 🔴 Unpredictable behavior (which implementation gets called depends on imports)

• 🔴 Code review nightmare (reviewing duplicate implementations wastes time)

The Cost of Code Duplication

This isn't just messy - it's expensive:

Real Example: A codebase with 800 functions had 52.7% duplication rate - 422 functions were duplicates. That's thousands of wasted lines of code.

How CodeWalker Solves This

CodeWalker indexes your codebase and lets Claude search before writing:

With CodeWalker

Day 1: You ask Claude to add CSV loading

Claude (internal): Let me check if CSV loading already exists...
> search_functions("load csv")

Found: load_csv_file() in src/data_loader.py

Claude: "I found an existing CSV loader. Let me use it instead of creating a new one."

Result:

# Claude imports existing function
from src.data_loader import load_csv_file

data = load_csv_file(path)

Day 5, 10, 15...: Same pattern - Claude finds and reuses existing code

Result after 2 weeks:

• ✅ 1 canonical CSV loading function (not 7)

• ✅ Consistent behavior across entire codebase

• ✅ Easy to maintain (fix bugs once, fixed everywhere)

• ✅ Predictable behavior (one implementation = one behavior)

• ✅ Fast code reviews (reviewers see reuse, not duplication)

Why This Problem Exists

LLMs Lack Architectural Awareness

Claude Code (and all LLMs) have a fundamental limitation:

❌ Can't see your codebase structure ❌ Can't search across files ❌ Can't remember what exists ❌ Can't detect duplicates

The technical reason: When Claude writes code, it only sees:

1. The current file you're editing

2. Recent conversation context

3. Maybe a few related files you showed it

What Claude DOESN'T see:

• That load_csv_file() already exists in src/data_loader.py

• That 3 other files have similar functions

• That your team has a canonical implementation

• Your codebase architecture and patterns

Result: Claude invents new implementations instead of reusing existing ones.

The "10 Developers, 0 Communication" Problem

Working with AI without CodeWalker is like having 10 developers who never talk to each other:

Developer 1 (Monday):    Creates load_csv_file()
Developer 2 (Tuesday):   Doesn't know about it, creates load_csv_data()
Developer 3 (Wednesday): Doesn't know about either, creates read_csv()
Developer 4 (Thursday):  Creates import_csv()
... and so on

Each "developer" (AI session) works in isolation, creating duplicates because they can't see what others did.

CodeWalker fixes this by giving AI a "shared memory" of your entire codebase.

Real-World Impact

Case Study: Elisity Project

Before CodeWalker:

• 800 total functions

• 422 duplicates (52.7% duplication rate)

• 33 direct pd.read_csv() calls (should use centralized loader)

• 11 duplicate print_summary() implementations

• 3 duplicate load_flow_data() functions with diverging behavior

With CodeWalker:

• Claude finds existing implementations before writing new code

• Duplication rate drops to near-zero for new code

• Codebase becomes more maintainable over time

Time Saved:

• Development: 30-40% less time rewriting existing code

• Code Review: Reviewers focus on new logic, not duplicate detection

• Bug Fixes: Fix once instead of hunting down 3-7 duplicates

How It Works

Architecture

┌─────────────────────┐
│   Your Codebase     │
│  (Python files)     │
└──────────┬──────────┘
           │
           │ AST Parser extracts
           │ function metadata
           ▼
┌─────────────────────┐
│   SQLite Index      │
│  (functions.db)     │
│                     │
│  • Function names   │
│  • Parameters       │
│  • Locations        │
│  • Docstrings       │
└──────────┬──────────┘
           │
           │ Claude queries via
           │ MCP protocol
           ▼
┌─────────────────────┐
│   Claude Code       │
│                     │
│  "Does load_csv     │
│   already exist?"   │
│                     │
│  → Yes! Use it      │
└─────────────────────┘

What Gets Indexed

For each function in your codebase:

• Name - load_csv_file

• Location - src/data_loader.py:42

• Parameters - (path, encoding='utf-8')

• Docstring - First line for quick understanding

• Type - Regular function, async function, or class method

• Decorators - @staticmethod, @cached, etc.

What's NOT stored: Function bodies, comments, string literals (only structural metadata).

Search Performance

• Parsing: ~100-200 files/second

• Indexing: ~1000 functions/second

• Search: Sub-millisecond SQLite queries

• Database size: ~1 KB per function

Example: 800 functions = ~800 KB database, indexed in < 5 seconds, searched in < 1ms.

Features

🔍 Search Before Writing

Tool: search_functions(query, exact=False)

Find existing functions before Claude writes new code:

> search_functions("load csv")

Found 3 functions:

• load_csv_file(path, encoding='utf-8')
  Location: src/data_loader.py:42
  Docs: Load CSV file with proper encoding handling

• FlowDataLoader.load_flows(flow_path, site_label)
  Location: modules/flow_loader.py:98
  Docs: Load flow data from CSV with site labeling

• read_raw_csv(filepath)
  Location: legacy/importer.py:156
  Docs: Legacy CSV reader (deprecated)

Claude sees these results and chooses to import the canonical implementation instead of creating a new one.

🔁 Detect Duplicates

Tool: find_duplicates()

Find functions with the same name in multiple files:

> find_duplicates()

⚠️  Found 3 function names with multiple implementations:

**load_flow_data** (3 implementations):
  - cohesion_analyzer.py:253
  - legacy/community_detector.py:440
  - policy_group_clustering.py:497

**format_bytes** (2 implementations):
  - utils.py:88
  - helpers.py:124

💡 Recommendation: Consolidate into single canonical implementations.

Use this to audit your codebase and identify consolidation opportunities.

🎯 Similar Signatures

Tool: find_similar_signatures(min_params=2)

Find functions with the same parameters (might be doing the same thing):

> find_similar_signatures(min_params=2)

Found 2 signature groups:

**Signature: (data, output_path)** - 4 functions:
  • save_to_csv in exporter.py:67
  • write_csv_file in writer.py:134
  • export_data in utils.py:203
  • save_results in analyzer.py:445

💡 These functions likely do the same thing with different names.

Catches semantic duplicates - functions that do the same thing but have different names.

📂 Multi-Project Support

Work on multiple projects without reconfiguring:

# One-time setup
> register_project("project-a", "/Users/jose/Projects/project-a")
> register_project("project-b", "/Users/jose/Projects/project-b")

# Daily use - auto-detects from your current directory
cd ~/Projects/project-a
> search_functions("auth")
[Auto-detected: project-a]
Found 5 functions...

cd ~/Projects/project-b
> search_functions("auth")
[Auto-detected: project-b]
Found 3 functions...

Features:

• ✅ Register unlimited projects

• ✅ Auto-detection from working directory

• ✅ Isolated indexes (no cross-contamination)

• ✅ Zero configuration switching

📊 Codebase Statistics

Tool: get_index_stats()

Understand your codebase at a glance:

> get_index_stats()

📊 CodeWalker Statistics:

Total Functions: 800
Total Files: 60
Unique Names: 765
Methods: 423
Async Functions: 67
Avg Parameters: 2.3

Duplication Rate: 4.4% (35 duplicates)
Last Indexed: 2026-03-18 10:35:00

Track duplication rate over time to measure improvement.

Quick Start

1. Install

git clone https://github.com/[username]/codewalker.git
cd codewalker
pip install -r requirements.txt

2. Configure Claude Code

Add to ~/.config/claude-code/mcp.json:

{
  "mcpServers": {
    "codewalker": {
      "command": "python3",
      "args": ["/absolute/path/to/codewalker/src/server.py"]
    }
  }
}

3. Register Your Projects

Restart Claude Code, then:

> register_project("my-project", "/absolute/path/to/your/project")

🔄 Registering project: my-project
📁 Path: /absolute/path/to/your/project

⏳ Indexing project...
Found 800 functions

✅ Indexing complete!

Total Functions: 800
Total Files: 60
Unique Names: 765

4. Start Using

CodeWalker now automatically prevents duplicate code:

You: "Add functionality to load CSV files"

Claude (internal):
  > search_functions("load csv")
  Found: load_csv_file() in src/data_loader.py

Claude: "I found an existing CSV loader at src/data_loader.py:42.
Let me use that instead of creating a new one:

from src.data_loader import load_csv_file
data = load_csv_file(path)

Available Tools

Project Management

• register_project(name, path) - Add a project to CodeWalker

• list_projects() - View all registered projects

• unregister_project(name) - Remove a project

• get_current_project() - Show which project is detected

Function Search

• search_functions(query, exact) - Find functions by name

• find_duplicates() - Detect duplicate function names

• find_similar_signatures(min_params) - Find functions with similar parameters

• get_file_functions(file_path) - List all functions in a file

• get_index_stats() - View codebase statistics

• reindex_repository() - Rebuild index after major changes

Use Cases

1. Prevent Duplication During Development

Before every implementation:

You: "Add user authentication"

Claude: Let me check if auth code already exists...
> search_functions("auth")
Found: authenticate_user() in src/auth.py

Claude: "I found existing auth code. Let me use it..."

2. Onboard to New Codebases

Explore unfamiliar code:

> search_functions("export")
Found 12 functions with "export" in the name

> get_file_functions("src/exporter.py")
Lists all 8 functions in the file with signatures and docs

Quickly understand what exists before writing new code.

3. Refactoring and Cleanup

Find consolidation opportunities:

> find_duplicates()
Found 15 duplicate function names

> find_similar_signatures()
Found 8 signature groups (functions with same params)

Systematically eliminate duplication.

4. Code Review

Reviewers can verify reuse:

Reviewer: "Why didn't you use the existing loader?"

Developer: "Let me check..."
> search_functions("load")
Found 3 loaders I didn't know about!

Catch missed reuse opportunities during review.

Comparison: With vs Without CodeWalker

Graph Theory Connection

CodeWalker treats your codebase as a graph:

• Vertices - Functions, classes, modules

• Edges - Imports, function calls, dependencies

• Walking - Traversing the graph to discover existing code

Graph concepts:

• Graph walk - Sequence of vertices (functions) and edges (calls)

• Traversal - Systematic exploration of the graph structure

• Random walks - Discovery algorithms (like PageRank)

• Tree walks - AST traversal (what the parser does)

This isn't just a metaphor - CodeWalker literally walks your Abstract Syntax Tree (AST) to build the function graph.

Roadmap & Future Development

CodeWalker v2.0.0 solves the core AI code duplication problem for Python projects. Future versions will add deeper analysis, broader language support, and smarter automation.

🔥 High Priority

Why these matter: These features provide immediate value for existing users and are most frequently requested.

• Incremental indexing - Currently, reindexing rebuilds the entire database. Incremental indexing would only update changed files, making reindexing 10-100x faster for large codebases. Impact: Seconds instead of minutes for 10k+ function codebases.

• Near-duplicate detection - Functions like load_csv, load_csv_data, and read_csv_file are semantically duplicates but have different names. Levenshtein distance matching would catch these "near-duplicates" that current exact/partial matching misses. Impact: Catch 20-30% more duplicates.

• Cross-project search - Search across all registered projects simultaneously. Useful for teams with shared utilities across multiple repos or monorepo users who want to find reusable code anywhere. Impact: Prevent reinventing wheels across project boundaries.

• Call graph analysis - Track what calls what to enable "blast radius" analysis ("what breaks if I change this function?") and identify unused code. Impact: Safer refactoring, dead code detection.

🎯 Medium Priority

Why these matter: These features enhance CodeWalker's intelligence and reduce manual effort.

• Semantic similarity (ML-based) - Detect functions that do the same thing with completely different names and signatures using embedding-based similarity. Example: save_to_csv(data, path) and export_results(df, filename) might be doing the same thing. Impact: Catch duplicates current signature matching misses.

• Auto-reindexing on file changes - Watch filesystem and automatically reindex when Python files change. No more manual reindex_repository() calls. Impact: Zero-maintenance index that's always current.

• Multi-language support - Extend beyond Python to JavaScript, TypeScript, Go, Rust, Java. Same duplication prevention for polyglot codebases. Impact: Unified duplication prevention across entire stack.

• Blast radius visualization - Show dependency trees and impact analysis when considering changes. "If I modify function X, these 15 functions are affected." Impact: Confident refactoring.

💡 Lower Priority

Why these matter: Nice-to-have features that improve developer experience but aren't critical to core functionality.

• Web UI - Visual interface for browsing functions, viewing call graphs, and exploring codebase structure in a browser. Alternative to CLI-only workflow. Impact: Better onboarding experience, visual learners benefit.

• VS Code extension - Native VS Code integration with inline suggestions ("⚠️ Similar function exists: use load_csv_file() instead"). Impact: Proactive duplicate prevention during typing.

• Import suggestions - When Claude is about to write new code, automatically suggest existing imports. "You're about to write X, but Y already exists - import it?" Impact: Even less manual searching.

• GitHub Action - CI/CD integration that fails PRs introducing duplicates above a threshold. Enforce duplication standards via automation. Impact: Prevent duplicates from ever being merged.

📊 Current Capabilities

What works today:

Language Support:

• ✅ Python (full support for functions, methods, async functions, decorators)

• 🚧 JavaScript, TypeScript, Go, Rust (on roadmap)

Analysis:

• ✅ Function names, signatures, locations, docstrings

• ✅ Parameter matching and signature comparison

• ✅ Duplicate detection (exact name matches)

• 🚧 Call graph analysis (planned)

• 🚧 Semantic similarity (planned)

• 🚧 Near-duplicate detection via Levenshtein distance (planned)

Indexing:

• ✅ Full repository indexing (~5 seconds for 800 functions)

• ✅ Manual reindexing on demand

• 🚧 Incremental updates (only changed files - planned)

• 🚧 Auto-reindexing on file changes (planned)

Search:

• ✅ Exact and partial name matching

• ✅ Parameter signature matching

• ✅ Multi-project support with auto-detection

• 🚧 Semantic search by behavior (planned)

• 🚧 Cross-project search (planned)

FAQ

Q: Does this work with other AI assistants?

Yes! CodeWalker uses the Model Context Protocol (MCP), which is an open standard. Any AI tool that supports MCP can use CodeWalker:

• Claude Code (tested)

• Claude Desktop (should work)

• Other MCP-compatible tools

Q: How much overhead does indexing add?

Very little:

• Initial indexing: ~5 seconds for 800 functions

• Reindexing: ~5 seconds (full rebuild)

• Search queries: < 1ms

• Memory: ~10 MB for typical projects

You barely notice it's there.

Q: What if my codebase is huge?

CodeWalker scales well:

• Tested on 800 functions / 60 files

• Should handle 10,000+ functions easily (SQLite scales)

• For massive codebases (100k+ functions), consider:

  • Incremental indexing (planned feature)

  • Multiple project registrations (already supported)

  • Excluding test files or generated code

Q: Can I use this on proprietary code?

Yes! Everything is local:

• ✅ Index stored locally (~/.codewalker)

• ✅ No data sent to external services

• ✅ No network requests during search

• ✅ Your code never leaves your machine

CodeWalker is 100% private.

Q: How is this different from IDE autocomplete?

Complementary, not competing:

IDE autocomplete:

• Works in single file

• Shows available imports

• Type-aware suggestions

• Real-time as you type

CodeWalker:

• Works across entire codebase

• Searches by semantic intent ("load csv")

• Finds duplicates proactively

• Used by AI during code generation

Use both - IDE for writing, CodeWalker for AI-assisted development.

Q: What about private/internal functions?

CodeWalker indexes everything:

• Public functions: ✅ Indexed

• Private functions (_private): ✅ Indexed

• Internal functions (__internal): ✅ Indexed

Why? Because you might want to reuse private functions too. Claude respects Python conventions (won't use _private from other modules without good reason), but knowing they exist prevents duplication.

Contributing

Contributions welcome! See CONTRIBUTING.md for guidelines.

Areas we need help:

• Multi-language support (JavaScript, TypeScript, Go)

• Incremental indexing

• Semantic similarity detection

• Performance optimization

License

MIT License - see LICENSE for details.

Free to use in personal and commercial projects.

Credits

Built to solve a real problem: Claude Code was creating duplicate implementations across a 60-file, 800-function codebase. CodeWalker eliminated the duplication.

Inspired by: Pharaoh (commercial tool for codebase intelligence)

Built with: Claude Sonnet 4.5 (dogfooding - using AI to build tools that improve AI)

Support

• Issues: GitHub Issues

• Discussions: GitHub Discussions

• Documentation: See guides in this repository

Summary

Problem: AI assistants can't see your codebase, causing massive code duplication.

Solution: CodeWalker indexes your codebase and lets AI search before writing.

Result: 40-60% reduction in duplicate code, faster development, cleaner codebase.

Get Started:

pip install -r requirements.txt
# Configure MCP (see Quick Start above)
> register_project("my-project", "/path/to/project")
> search_functions("whatever you're about to write")

Stop duplicating code. Start walking your codebase. 🚀