ohh-my-excel 0.1.0

Persistent, incrementally-updated knowledge graph for Excel workbooks — safe, fast, token-efficient MCP server

Ohh-my-excel

Safe · Fast · Token-Efficient — Knowledge Graph for Excel Workbooks

Existing Excel MCP servers are CRUD-only — they read/write cells and dump raw data. None provide structural understanding of formula dependencies, blast radius analysis, or token-efficient context. When an AI assistant opens a 50-sheet financial model, it reads everything, burning tokens and missing critical relationships.

Ohh-my-excel fixes that. It parses Excel workbooks into a structural knowledge graph of sheets, tables, cells, formulas, and dependencies — then exposes it via MCP tools so AI assistants get precise, token-efficient context instead of dumping entire spreadsheets.

---

Quick Start

pip install ohh-my-excel
ohh-my-excel build path/to/workbook.xlsx

Then open your AI assistant and ask:

Build the Excel knowledge graph for this workbook

The initial build takes ~10 seconds for a 100-sheet workbook. After that, incremental updates complete in under 500ms.

ohh-my-excel build path/to/workbook.xlsx    # Build or rebuild graph
ohh-my-excel update path/to/workbook.xlsx   # Incremental update
ohh-my-excel status path/to/workbook.xlsx   # Show graph stats
ohh-my-excel serve                          # Start MCP server

Requires Python 3.10+.

---

How It Works

┌─────────────────────────────────────────────────────────┐
│                    MCP Client (AI Agent)                  │
│         OpenCode / Claude Code / Codex / Cursor          │
└────────────────────┬────────────────────────────────────┘
                     │  stdio / Streamable HTTP
                     ▼
┌─────────────────────────────────────────────────────────┐
│              FastMCP Server (38 tools + 5 prompts)       │
└────────────────────┬────────────────────────────────────┘
                     │
                     ▼
┌─────────────────────────────────────────────────────────┐
│              SQLite Graph Store (WAL mode)               │
│         nodes + edges + FTS5 + NetworkX cache           │
│         .excel-graph/graph.db                           │
└────────────────────┬────────────────────────────────────┘
                     │
                     ▼
┌─────────────────────────────────────────────────────────┐
│              Excel Parser + Formula AST                  │
│         openpyxl + formulas + cell ref extraction       │
└────────────────────┬────────────────────────────────────┘
                     │
                     ▼
┌─────────────────────────────────────────────────────────┐
│              Excel Workbook (.xlsx / .xlsm)              │
└─────────────────────────────────────────────────────────┘

Blast-radius analysis

When a cell changes, the graph traces every dependent formula, cross-sheet reference, and data flow that could be affected. Your AI reads only these instead of scanning the whole workbook.

Incremental updates in < 500ms

On every file save, a SHA-256 hash comparison detects changes. Only changed sheets are re-parsed. A 100-sheet workbook re-indexes relevant cells in under 500ms.

Token efficiency: 10-250x reduction

Scenario	Without Graph	With Graph	Reduction
Audit 50-sheet workbook	~500k tokens	~50k tokens	~10x
Review cell change	~500k tokens	~5k tokens	~100x
Find formula dependencies	~500k tokens	~2k tokens	~250x
Understand data flow	~500k tokens	~10k tokens	~50x

---

38 MCP Tools

Category	Tools	Description
Graph Setup	3	Build, update, post-process, stats
Exploration	4	Minimal context, search, query, traverse
Dependency & Impact	3	Blast radius, change detection, affected flows
Formula Analysis	5	Dependencies, dependents, circular refs, broken refs, complexity
Sheet & Workbook	5	List sheets, sheet info, cross-sheet refs, tables, named ranges
Data Flow	2	List flows, get flow details
Architecture & Quality	6	Communities, hub cells, bridge cells, knowledge gaps, questions, architecture overview
Visualization & Export	2	Interactive HTML graph, JSON/CSV/GraphML export
Workbook Generation	8	Create from prompt, from data, from template, add sheet, add chart, format, formulas, validate

Key Tools

Tool	What it does
build_or_update_graph_tool	Build or incrementally update the Excel knowledge graph
get_minimal_context_tool	Ultra-compact ~100 token workbook summary — call this first
get_impact_radius_tool	Blast radius when a cell/formula changes
query_graph_tool	Precedents, dependents, cross-sheet refs, contains
traverse_graph_tool	BFS/DFS from any cell/sheet with token budget
find_circular_references_tool	Detect circular formula dependencies
detect_changes_tool	Risk-scored change analysis between versions
get_architecture_overview_tool	High-level workbook structure from sheet communities
generate_workbook_tool	Create workbook from natural language prompt
visualize_graph_tool	Generate interactive HTML dependency graph

MCP Prompts (5)

audit_workbook, debug_formula, data_flow_map, onboard_analyst, pre_merge_check

---

Three Pillars

Safe

• Automatic backups before every destructive write (.excel-graph/backups/)
• Formula validation via AST parsing before writing
• Circular reference detection via NetworkX before saving
• No eval/exec — zero arbitrary code execution
• Path traversal prevention — _validate_file_path() sanitizes all input
• Max file size limits — configurable (default: 100MB read, 50MB generate)
• Backup retention — 7 days by default, max 10 per workbook

Fast

• SQLite WAL mode — concurrent reads, zero lock contention
• SHA-256 hash skip — only re-parse changed sheets
• Range node optimization — SUM(A1:A100) = 1 edge, not 100
• Lazy cell loading — values loaded on query, not during build
• Batch cell reads — entire sheet ranges in one call
• Cached NetworkX graph — O(1) query response after build
• Async thread offload — no event loop blocking

Token-Efficient

• get_minimal_context_tool — ~100 tokens: stats + next_tool_suggestions
• detail_level="minimal" — all tools support compact output
• Structural context over raw data — "Cell C5 depends on A1:A100 via SUM" instead of 100 cell values
• next_tool_suggestions — every response tells AI the optimal next tool
• Target: ≤5 tool calls per task, ≤800 total tokens

---

Architecture

excel_graph_mcp/
├── __init__.py
├── constants.py       # Config, schema, performance targets, env vars
├── formula_parser.py  # Formula AST → cell/range refs, functions, nesting depth
├── parser.py          # openpyxl multi-sheet parser (cells, formulas, tables, charts)
├── graph.py           # SQLite WAL graph store (nodes, edges, FTS5, BFS, NetworkX cache)
├── dependency.py      # Dependency graph builder, circular/broken ref detection
├── flows.py           # Data flow detection (input → calculation → output chains)
├── communities.py     # Sheet community detection + architecture overview
├── changes.py         # Risk-scored version-to-version change analysis
├── analysis.py        # Cross-sheet refs, hub/bridge cells, knowledge gaps, suggested questions
├── exports.py         # JSON/CSV/GraphML export + Pyvis/Matplotlib visualization
├── incremental.py     # SHA-256 hash change detection, auto-update
├── generation.py      # 8 workbook creation/modification/validation tools
├── main.py            # FastMCP server — 38 tools + 5 prompts
├── cli.py             # CLI: build, update, status, serve, version
└── tools/             # Tool implementations (build, query, context)

Graph Schema

Node types: Workbook, Sheet, Cell, Formula, Range, Table, NamedRange, Chart

Edge types: CONTAINS, REFERENCES, RANGE_REF, DEPENDS_ON, CROSS_SHEET_REF, TABLE_REF, CHART_SOURCE, NAMED_REF

Confidence levels: EXTRACTED (AST), INFERRED (range expansion), AMBIGUOUS (INDIRECT/OFFSET)

---

Configuration

Environment Variables

Variable	Description	Default
EXCEL_GRAPH_BACKUP_RETENTION_DAYS	Days to keep backups	7
EXCEL_GRAPH_MAX_BACKUPS	Max backups per workbook	10
EXCEL_GRAPH_WATCH_DEBOUNCE_MS	Watch mode debounce	5000
EXCEL_GRAPH_RATE_LIMIT	Generation requests/min	10
EXCEL_GRAPH_MAX_FILE_SIZE_MB	Max read file size	100
EXCEL_GRAPH_GENERATE_MAX_SIZE_MB	Max generate file size	50

Optional Dependency Groups

pip install ohh-my-excel[embeddings]     # Vector embeddings (sentence-transformers)
pip install ohh-my-excel[visualization]  # Pyvis + Plotly visualization
pip install ohh-my-excel[dev]            # Development (pytest, ruff, bandit)

---

Supported Formats

Format	Read	Write	Formula Parsing	Notes
.xlsx	✅	✅	✅	Primary format
.xlsm	✅	✅	✅	+ VBA macro extraction
.xltx	✅	✅	✅	Templates
.xltm	✅	✅	✅	Macro-enabled templates
.xlsb	✅	❌	⚠️	Binary format (via pyxlsb)
.xls	⚠️	❌	❌	Legacy (via xlrd)
.csv	✅	✅	❌	No formulas

---

Performance Targets

Metric	Target
Initial build (100-sheet workbook)	<30s
Incremental update (1 cell change)	<500ms
Query response (any tool)	<100ms P95
Search response	<10ms
Memory usage (100-sheet workbook)	<200MB
Graph DB size (100-sheet workbook)	<50MB

---

Development

git clone https://github.com/adityaidev/Ohh-my-excel.git
cd Ohh-my-excel
pip install -e ".[dev]"
pytest

Running CI Checks Locally

pytest tests/ --tb=short -q                    # Run tests
ruff check excel_graph_mcp/                     # Lint
bandit -r excel_graph_mcp/ -c pyproject.toml    # Security scan

---

What's Next

Feature	Status
HyperFormula sidecar — dual formula engine (JS + Python)	Planned
VBA analysis tools — extract, summarize, explain VBA macros	Planned
Template expansion — 20+ built-in templates across 6 categories	Planned
Embedding-based semantic search — vector search via sentence-transformers	Planned
Power Query / DAX roadmap — financial analyst integration	Future
Excel Online / Graph API — read from SharePoint/OneDrive	Future
Claude in Excel differentiation — MCP-native, any-AI-client support	Ongoing

---

Licence

MIT. See LICENSE.

github.com/adityaidev/Ohh-my-excel

pip install ohh-my-excel && ohh-my-excel build workbook.xlsx
_{38 MCP Tools · 5 Prompts · 62 Tests · Safe · Fast · Token-Efficient}