GraphHub

Quick Start · How It Works · MCP Tools · Token Savings · Roadmap

The Problem

AI coding assistants lose context between sessions. They re-read the same files, re-learn your codebase, and burn tokens on questions they've already answered.

94% fewer tokens. Same answer. Every time.

⚡ Quick Start

npm install graphhub

One command. GraphHub auto-detects every AI agent on your machine and writes to their global configs — no per-project setup, ever.

npx graphhub setup                                # auto-detect + install present clients
npx graphhub setup --force                        # install for all 5 clients regardless
npx graphhub setup --client claude-code,kilo-cli  # pick specific clients
npx graphhub setup --dry-run                      # preview what would be installed
npx graphhub setup --list                         # list all supported clients
npx graphhub uninstall-all                        # remove from all clients

git clone https://github.com/slnquangtran/Graph-Hub.git
cd Graph-Hub && npm install

npm run index -- /path/to/your/project
npm run setup -- /path/to/your/project

🧠 How It Works

  Your Source Code
        │
        ▼
  ┌─────────────────┐     ┌──────────────────┐
  │  Tree-sitter    │────▶│  KuzuDB Graph DB │
  │  AST Parser     │     │                  │
  │  TS · JS · PY   │     │  File ──CONTAINS─▶ Symbol
  └─────────────────┘     │  Symbol ──CALLS──▶ Symbol
        │                 │  File ──IMPORTS──▶ File
        ▼                 │  Symbol ──INHERITS▶ Symbol
  ┌─────────────────┐     └────────┬─────────┘
  │  Local Embeddings│             │
  │  (MiniLM-L6-v2) │      ┌──────▼──────────┐
  │  No API needed   │      │   MCP Server    │
  └─────────────────┘      │   REST API :9000 │
                            │   Dashboard :5173│
                            └─────────────────┘

Supported AI Agents

📊 Token Savings

Real measurements on GraphHub's own codebase — not synthetic benchmarks.

Result: 5× more tasks in the same context window.

🛠 MCP Tools

GraphHub registers 27 MCP tools across 7 categories. Click each to expand.

// Natural language → code
semantic_search({ query: "authentication validation", mode: "hybrid" })

// Exact or fuzzy symbol name
search_by_name({ name: "validateToken" })

// Results grouped by file
search_grouped({ query: "error handling middleware" })

// Cosine-similar symbols
find_similar({ name: "parseRequest", top_k: 5 })

// Explain which strategy was chosen
explain_search({ query: "jwt token refresh" })

// Callers + callees of a symbol
get_context({ name: "validateToken" })

// Full blast radius before you edit
impact_analysis({ name: "handleRequest" })

// Callers/callees for many symbols at once
batch_context({ names: ["validateToken", "handleRequest", "parseBody"] })

// Symbols in git-changed files + risk buckets
changed_symbols({ diff: "staged" })

// Raw Cypher for power users
query_graph({ cypher: "MATCH (s:Symbol)-[:CALLS]->(t) RETURN s.name, t.name LIMIT 20" })

// All symbols in a file
get_file_symbols({ path: "src/services/auth/token.ts" })

// RAG search + context + impact + next steps in one call
debug_trace({ query: "null pointer in auth middleware", top_k: 3 })

// Find functions/classes nobody calls — safe to delete?
find_dead_code({ kinds: ["function", "method"] })

// Find near-duplicate implementations (uses stored embeddings, no API cost)
find_duplicates({ name: "validateToken", min_similarity: 0.85, cross_file_only: true })

// Detect circular import chains and mutual-recursion call cycles
find_cycles({ type: "both", max_length: 3 })

// Save a learning
remember({ content: "Auth uses JWT with 15m TTL", type: "learning", project: "myapp" })

// Retrieve by similarity
recall({ query: "how does auth work?" })

// Chronological view
timeline({ limit: 20, project: "myapp" })

// Update or delete
update_observation({ id: "abc-123", content: "Updated: now uses refresh tokens" })
forget({ project: "myapp", type: "learning" })

// Linked to specific symbols
related_observations({ symbol: "validateToken" })

// Store a bug fix pattern
remember_bugfix({
  symptom: "TypeError: Cannot read property user of undefined",
  root_cause: "req.session was null on the /guest route",
  fix: "Added session guard at middleware entry",
})

// Recall by symptom similarity — even next week
recall_bugfix({ symptom: "undefined user property on guest route" })

// Cache which skill worked for a task
remember_skill_choice({
  task_description: "rename a function safely across the repo",
  skill_path: ".claude/skills/refactoring/SKILL.md",
  outcome: "success",
})

// Route future similar tasks to the same skill
recall_skill_choice({ task_description: "rename a symbol in multiple files" })

✨ Features

🧩 Core Intelligence

• Knowledge Graph — functions, classes, imports, call chains stored in KuzuDB

• Semantic Search — natural language queries, no API costs (local MiniLM-L6-v2)

• Impact Analysis — see blast radius before editing, direct + indirect callers

• Session Memory — remember / recall / forget across sessions

• Pattern Memory — recall past bug fixes and skill choices by similarity

• One-Shot Debug — debug_trace chains search → context → impact in one call

• Batch Context — callers/callees for many symbols in one round trip

🚀 Developer Experience

• One Command — npm install graphhub configures all detected agents globally

• 5 Agents — Claude Code, OpenCode, Gemini CLI, Antigravity, Kilo CLI

• Auto-Reindex — PostToolUse hook keeps graph fresh after commits

• Always-On — PreToolUse hook reminds Claude about the graph before every search

• Graph Report — auto-generated GRAPH_REPORT.md with god nodes + clusters

• 100% Local — all data stays in .graphhub/, no telemetry, no cloud

• React Dashboard — interactive call graph at :5173

🌐 Language Support

⚙️ Configuration

npx graphhub setup

Auto-detects and writes to each agent's global config — no per-project setup needed.

Existing keys are preserved — only the graphhub entry is added or updated. Use --force to install for agents that aren't yet present. A postinstall hook runs automatically after npm install graphhub; opt out with GRAPHHUB_NO_INSTALL=1, CI=1, or npm_config_global=true.

npm run install-claude

Adds PreToolUse/PostToolUse hooks and updates CLAUDE.md. The setup command above covers MCP config only.

{
  "mcpServers": {
    "graphhub": {
      "command": "npx",
      "args": ["tsx", "src/index.ts", "serve"],
      "cwd": "/path/to/graphhub"
    }
  }
}

📋 Commands

🏗 Tech Stack

🗺 Roadmap

• Session memory (remember / recall / forget)

• Always-on PreToolUse hooks

• PostToolUse auto-reindex

• Graph report generation

• Verified 85% token reduction

• Class hierarchy edges (INHERITS, IMPLEMENTS)

• One-shot debug_trace and bulk batch_context

• Pattern memory for bug fixes and skill routing

• One-command multi-client setup (Claude Code, OpenCode, Gemini CLI, Antigravity, Kilo CLI)

• Global install — writes to home-dir configs so all projects benefit without per-project setup

• .gitignore support (cross-platform, including Windows path normalization)

• Proper database lifecycle — close() releases the file lock on MCP server, watch mode, and CLI exit

Coming next:

• Dead code detection (find_dead_code)

• Duplicate function detection (find_duplicates)

• Circular dependency / call cycle detection (find_cycles)

• Worker thread indexing for large repos

• Native Go / Rust / Java Tree-sitter grammars

• Community detection (Leiden algorithm)

📜 Changelog

• find_dead_code — Find functions, methods, and classes with zero callers. Common entry-point patterns (main, init, onX, handleX…) are filtered automatically; pass include_entry_points: true to see everything.

• find_duplicates — Find near-duplicate implementations of any symbol using stored cosine embeddings. No API cost — works offline. Supports min_similarity threshold and cross_file_only filter.

• find_cycles — Detect circular import chains and mutual-recursion call cycles up to length 3. Reports each cycle as an ordered list of file paths (import) or symbol names (call).

• 14 new tests — Full coverage of entry-point filtering, cycle detection, similarity threshold, and cross-file filtering.

• Global install — All adapters now write to home-dir global configs (~/.claude/, ~/.gemini/, etc.) so one npx graphhub setup --force covers every project without per-project setup

• Kilo CLI support — New adapter for Kilo CLI (~/.config/kilo/kilo.json)

• graphhubDir fix — npx graphhub setup now correctly resolves the package location from import.meta.url instead of process.cwd() (which pointed to the user's project, causing broken MCP server paths)

• Gemini CLI fix — Was writing to <project>/.gemini/ instead of ~/.gemini/; fixed to use home dir

• Windows path fix — MCP server entry paths now always use forward slashes in JSON configs

• DB lifecycle fix — MCP server, watch mode, and all CLI commands now properly release the KuzuDB file lock on exit

• .gitignore support — cross-platform path normalization (fixes Windows \ vs / mismatch)

• Python parser — upgraded to full Tree-sitter AST extraction (functions, classes, calls, inheritance)

• Import alias resolution — import { foo as bar } now stores foo as the specifier, fixing cross-file call graph edges

• Fuzzy symbol search — case-insensitive (toLower(s.name) CONTAINS query)

• forget() safety — requires at least one filter; no longer silently deletes all observations

• MCP get_file_symbols — fixed: was returning an empty object (raw DB cursor)

• Error resilience — per-file parse errors no longer abort the entire indexing run

• BigInt coercion — KuzuDB count() results coerced to Number; JSON.stringify no longer throws on stats

🤝 Contributing

Contributions are welcome! Feel free to open an issue or submit a PR.

git clone https://github.com/slnquangtran/Graph-Hub.git
cd Graph-Hub && npm install

npm test                    # run 77 tests
npm run index -- ./src      # index GraphHub itself

🔒 Privacy

📄 License

ISC © 2024

If GraphHub saves you tokens, a ⭐ star helps others find it.

Report Bug · Request Feature · Discussions