graphfocus 0.2.0

Turn any folder of code into a queryable knowledge graph. Multi-language AST extraction with plugin architecture and MCP server for AI tools.

GraphFocus

Turn any folder of code into a queryable knowledge graph. Multi-language AST extraction with plugin architecture, designed to feed AI coding tools (Claude Desktop, Trae AI, Cursor, Windsurf) with focused, token-efficient context via MCP.

Why

Most code intelligence tools either parse only one language (so a fullstack Spring Boot + Vue project becomes a black box) or shovel raw files to an LLM until you blow your context window. GraphFocus extracts a deterministic knowledge graph via tree-sitter ASTs across 19 languages, links cross-language references (e.g. @Entity OrderEntity ↔ SQL orders table), and exposes it to AI tools so the model can ask find_callers("validate") instead of grepping the whole repo.

Features

• 🌳 19 languages with dedicated extractors — Python, Java, TypeScript/JavaScript/React, Vue, C#, Go, Kotlin, Rust, Swift, Ruby, PHP, Scala, Lua, C, C++, Dart, R, SQL, PL/SQL
• 🔌 Plugin architecture — Add new languages by implementing a single interface
• 🔗 Cross-language linker — Java/C# @Entity/@Table automatically wired to SQL/PL-SQL tables
• 🤖 MCP server — Plug into any MCP-compatible AI tool (Trae, Cursor, Claude Desktop…) so the LLM queries the graph directly
• 📁 Multiple outputs — Interactive HTML viz, JSON, Markdown report, Obsidian vault, AI_SUMMARY.md (dense LLM context)
• ⚡ Smart caching — SQLite-based, only re-processes changed files
• 🚀 FastAPI server — REST endpoints to query the graph from any tool
• 🧠 Optional LLM enrichment — Use Claude/GPT to extract semantic relationships from docs and images

Install

Requires: Python 3.11+

# Core + all 19 extractors
pip install graphfocus

# Add the MCP server for AI tools
pip install "graphfocus[ai]"

# Add the FastAPI server
pip install "graphfocus[api]"

# Everything
pip install "graphfocus[all]"

Quick start

# Analyze the current directory; --update enables incremental cache;
# --ai writes the dense LLM summary; --obsidian writes a navigable vault.
graphfocus analyze . --update --ai --obsidian

# Output lands in ./graphfocus-out/:
#   graph.json        ← raw data
#   graph.html        ← interactive D3 visualization
#   GRAPH_REPORT.md   ← human report
#   AI_SUMMARY.md     ← LLM-friendly dense map (paste into a chat)
#   obsidian/         ← one .md per node, with wikilinks
#   .cache.db         ← incremental cache

Query the graph from the CLI

graphfocus find UserService                  # search by label or id
graphfocus neighbors userservice_user 2      # explore neighborhood
graphfocus callers validate                  # who calls validate()?
graphfocus languages                         # list active extractors

Plug into your AI tool (MCP)

GraphFocus ships an MCP server that exposes the graph to any compatible AI tool. The LLM gains these tools:

Tool	What it does
find_symbol	Search nodes by label/id, filter by language or kind
get_node	Full info on one node + its incoming/outgoing edges
get_neighbors	Walk N hops out from a node
find_callers	Who calls this function/method
find_path	Shortest path between two nodes
list_languages	What's in the graph
get_stats	Counts by kind and relation
cross_language_links	Only edges that cross languages

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "graphfocus": {
      "command": "graphfocus",
      "args": ["mcp"],
      "cwd": "/absolute/path/to/your/project"
    }
  }
}

Trae AI / Cursor / Windsurf / Continue.dev

In the tool's MCP config (.cursor/mcp.json, Trae settings, etc.):

{
  "mcpServers": {
    "graphfocus": {
      "command": "graphfocus",
      "args": ["mcp"]
    }
  }
}

Restart the AI tool and ask things like "who calls validateUser?" — the model invokes find_callers and gets back a 200-byte JSON answer instead of grepping the whole codebase.

HTTP transport

For tools that prefer HTTP+SSE over stdio:

graphfocus mcp --http --port 8765

Token efficiency comparison

For a mid-sized Spring Boot project:

Approach	Tokens sent to LLM
Raw source files	30–100K
Full graph.json	15–40K
AI_SUMMARY.md + a few Obsidian notes	1–3K
MCP find_callers("X") round-trip	200–800 per question

Supported languages

Language	Parser	Extracts
Python	tree-sitter	classes, functions, imports, calls, inheritance
Java	tree-sitter	classes, interfaces, methods, Spring annotations
TypeScript / JavaScript / React	tree-sitter (tsx)	classes, interfaces, type aliases, components
Vue	regex + delegated TS parser	component node, <script> symbols, child component refs in <template>
C#	tree-sitter	classes, methods, namespaces, attributes
Go	tree-sitter	structs, interfaces, methods (attached to receiver)
Kotlin	tree-sitter	classes, data classes, objects, Spring annotations
Rust	tree-sitter	structs, enums, traits, impl Trait for Type
Swift	tree-sitter	protocols, structs, classes, methods
Ruby	tree-sitter	modules, classes, methods, requires
PHP	tree-sitter	namespaces, interfaces, classes, methods
Scala	tree-sitter	traits, classes, case classes, objects
Lua	tree-sitter	functions, module methods (M.foo), requires
C	tree-sitter	structs, typedefs, functions, includes
C++	tree-sitter	namespaces, classes, out-of-class methods
Dart	regex	classes, mixins, abstract classes, methods
R	regex	functions, library/source imports
SQL	regex (dialect-agnostic)	tables, columns, FKs, views, indexes
PL/SQL	regex	packages, procedures, functions, triggers

Architecture

Input files
   ↓
Detector  →  Registry  →  19 Extractors (per language)
                              ↓
                       Merger (dedup nodes)
                              ↓
                       Cross-language linker
                       (Java @Entity ↔ SQL table)
                              ↓
                Outputs:
                  • graph.json
                  • graph.html (D3 force-directed)
                  • GRAPH_REPORT.md
                  • AI_SUMMARY.md (dense for LLM)
                  • obsidian/  (one .md per node)
                              ↓
                MCP server / FastAPI / Query CLI
                              ↓
                Any AI tool over MCP

Adding a new language

from graphfocus.extractors.base import LanguageExtractor, ExtractionResult

class MyLanguageExtractor(LanguageExtractor):
    @property
    def language_name(self) -> str:
        return "mylanguage"

    @property
    def extensions(self) -> set[str]:
        return {".ml", ".mli"}

    def extract(self, path: Path) -> ExtractionResult:
        # Your extraction logic using tree-sitter or regex
        ...

Register the class in graphfocus/extractors/registry.py and add a test + fixture. See CONVENTIONS.md for details.

Development

make dev      # install with all extras
make test     # run pytest
make lint     # ruff check
make format   # ruff format
make serve    # start the FastAPI server

Release

Publishing to PyPI is automated via GitHub Actions and PyPI Trusted Publishing — no API tokens are stored in this repo.

One-time setup on pypi.org (project owner only):

1. Log in to https://pypi.org and go to the project (after the first manual upload, or use a pending publisher).
2. Manage → Publishing → Add a new publisher → fill in:
   • Owner: bamc300
   • Repository name: graphfocus
   • Workflow name: publish.yml
   • Environment name: pypi

Each release:

# 1. bump the version in pyproject.toml
# 2. commit + tag
git commit -am "chore: bump version to 0.2.0"
git tag v0.2.0
git push && git push --tags
# 3. the publish.yml workflow runs automatically and uploads to PyPI

License

MIT