ModelForge

Bulge-tier Excel financial model factory for credit & structured finance. Every cell live-formulated. Every number traceable back to the source document page it came from.

A developer tool for analysts and engineers who build credit and corporate-finance models programmatically. Covers unitranche, sponsor-backed LBO, project finance, real estate credit, NPL, structured credit, restructuring, M&A, DCF and IPO templates. Extensible to any asset class.

The moat: builds are byte-identical deterministic (same spec → same workbook bytes, every run) and ship with a verifiable manifest + certificate — formula integrity, accounting/conservation invariants (balance sheet balances, cash ties out), and SHA-256 hashes of spec + sources + workbook. Run certify --strict / build --trust-strict and it's fail-closed: non-zero exit on any integrity violation, so a broken model never ships. That's model generation with a portable audit trail — not just generation. For an AI agent or an app that emits financial models, it's the layer that turns "the LLM produced a spreadsheet" into "here is a certificate that the spreadsheet is internally correct and reproducible."

🚀 Using ModelForge in production — or want managed features, priority support, or a specific template/connector? Tell me about your use case → — I read every one.

What this solves

• Your agent needs to produce an Excel model from a structured spec — without an LLM hallucinating numbers directly into cells. ModelForge keeps the model deterministic: the LLM writes a typed YAML spec with source IDs, and a Python builder emits the live-formula workbook.
• You need every output number to be auditable back to where it came from — without manually maintaining a sources sheet. Each hardcoded input carries a source ID, and the model's linkage graph is persisted to SQLite so a cell can be traced to its driver, source, and document page.
• You want a model that recalculates instead of being a static dump — without writing formula strings by hand. Every cell is a real Excel formula, with named ranges, sign conventions, and WORST/BASE/BEST scenario toggles wired across sheets.
• You need to gate a workbook for review — without eyeballing it. The QC tool runs an automated structural check suite (QC sheet present, named ranges populated, source references resolve, print areas set, no orphan sheets) and returns a per-check pass/fail report.
• You need to triage many candidate deals fast — without building a workbook for each one. The screening tool filters and ranks a directory of spec YAMLs by quantitative criteria (margins, leverage, IRR) on their screening: block alone.
• You want the whole pipeline available to an AI assistant — without bespoke glue code. ModelForge ships an MCP server (modelforge-mcp) so agents in Claude Code, Cursor, Cline, or ChatGPT Enterprise can list templates, build, QC, trace lineage, ingest a data room, and export deliverables.

Use it inside Claude Code, Cursor, ChatGPT Enterprise (MCP-native)

PyPI name: modelforge-finance (the unscoped modelforge was taken by source{d}'s ML library). Import name stays modelforge.

pip install "modelforge-finance[mcp,export]"

# wire into your MCP client config:
{
  "mcpServers": {
    "modelforge": { "command": "modelforge-mcp" }
  }
}

Then in your AI assistant:

"Build me a unitranche LBO model from this YAML spec, export the committee deck."

Tools available: list_templates · build_model · qc_workbook · list_sources · lineage_walk · ingest_dataroom · screen_deals · compute_tax · export_pptx · export_docx · plus 7 unified-feed tools (data_providers_status · quote · history · fundamentals · search_filings · entity_lookup · search_securities) across a 14-provider data stack.

The architectural principle

LLMs produce specs + sources + narrative. Deterministic Python produces the workbook.

The LLM never writes a number into a cell. It writes a typed YAML spec with source IDs. A deterministic builder emits the Excel via openpyxl. A QC gate validates before export. Excel is a render of a linkage graph; the graph is persisted to SQLite and is the canonical artifact.

Quality standards (bulge-tier, non-negotiable)

Formatting

• Blue = hardcoded input. Black = formula. Green = cross-sheet link. Red = warning.
• No mixed formulas (no magic numbers embedded). Named ranges for every driver.
• Costs NEGATIVE (sign convention enforced and checked).
• EN primary labels, multi-language secondary (DE / ES / IT shipped; SV / NO / DA / NL on the v0.10 roadmap as design-partner asks).
• Historical vs Projected column separator, obvious.
• Check row at top of every sheet (BS balance, CFS tie, covenant headroom — TRUE or 0).

Sourcing

• Every hardcoded cell has a comment with source ID (S-001, S-002, ...).
• Sources sheet lists each source: doc, page, publisher, date, URL, verified-flag.
• Assumptions (not sourced) tagged A-001 with rationale + confidence H/M/L.

Scenarios

• WORST / BASE / BEST toggle on Assumptions. Drives every sheet via CHOOSE.
• Every sheet respects the toggle — no orphan assumptions.

Audit

• QC sheet with 8 automated checks, all must pass.
• Revision log on Cover.
• Named ranges mandatory.
• Print areas set. Print-ready on every sheet.

Quick start

pip install "modelforge-finance[mcp,export]"

# Scaffold a ready-to-build spec — no repo checkout needed (works for any of the 19
# templates; run `modelforge list-templates` to see them all)
modelforge scaffold dcf -o demo_dcf.yaml

# Build it: live-formula workbook + linkage graph + manifest sidecar
modelforge build demo_dcf.yaml            # -> output/demo_dcf.xlsx

# Certify the delivered artifact: zero formula errors, byte-identical, manifest-valid
modelforge certify output/demo_dcf.xlsx

Trust Layer v1 (new in v0.9.7)

Why should a buyer trust the number in cell B42?

The Trust Layer is a semantic gate (separate from the structural QC gate). It answers the question every IC asks in the first five minutes: is this number plausible? It catches issues like a DCF EV that's 8× the company's real market cap before the model ever leaves QA.

25+ built-in rules cover all shipped templates:

• DCF: WACC band (3-25%), terminal growth ≤ GDP + 1%, EV vs market-cap deviation, terminal-value share, sensitivity-table monotonicity
• Three-statement: balance-sheet integrity, cash reconciliation, retained-earnings link
• NPL: cumulative recovery ≤ 100%, vintage staircase monotone
• Project finance: DSCR floor, wire degradation > 0, P90 < P50
• Sponsor LBO: XIRR plausibility, multiple expansion vs entry
• M&A / fairness / structured credit / unitranche / credit memo: per-template plausibility

Each violation produces a RedFlags worksheet inside the built workbook with severity (info / warn / fail), the rule that fired, expected-vs-actual, and the recommended remediation.

modelforge audit-all examples/   # every shipped example, 0 FAIL violations in current ship

See AUDIT_REPORT.md for the current ship's audit.

Data-room ingestion (v0.3.1)

Turn a directory of PDFs, XLSXs and CSVs into a validated ModelForge YAML spec using Claude Opus. Every extracted number traces back to a doc page via the auto-built Sources registry.

pip install -e .[ingest]                # installs anthropic, pdfplumber, pypdf
export ANTHROPIC_API_KEY=sk-ant-...      # required

modelforge ingest path/to/dataroom/ \
    --template project_finance \
    -o output/my_deal.yaml --verbose

# Review output/my_deal.yaml + output/my_deal.ingestion.md
# (INGESTION_REPORT.md lists every extracted field, S-id, confidence)

modelforge build output/my_deal.yaml     # produces the workbook
modelforge qc output/my_deal.xlsx        # 8/8 quality gate

Supported template: project_finance (MVP). Templates 1, 3, 5-8 queued for v0.3.2.

Package layout

modelforge/
├── graph/            # First-class linkage graph (nodes, edges, SQLite persistence)
├── spec/             # Pydantic schemas per template
│   ├── base.py       # Source, Assumption, Scenario, Target (shared types)
│   └── unitranche.py # Template 1: Unitranche LBO
├── builder/          # Deterministic openpyxl writer
│   ├── styles.py     # Bulge-tier formatting library
│   ├── formulas.py   # Formula string builders
│   ├── i18n.py       # EN/IT label dictionary
│   ├── workbook.py   # Top-level builder
│   └── sheets/       # One module per sheet (cover, sources, assumptions, ...)
├── qc/               # Quality gate (8 structural checks + PDF report)
├── data/             # Market data loaders (Damodaran, ECB, Borsa minibond)
└── cli.py            # build | certify | qc | scaffold | validate | screen | ingest | ...

Templates (19: 17 shipped + 2 preview)