English · Italiano
Your agent reads the documents and writes the report. Fairmark does the math —
exactly, reproducibly, with every step on the record.
Documentation · Quickstart · Methods · Roadmap · Contributing
Language models are excellent at reading financial statements and drafting an investment memo. They are not reliable calculators: a discount factor rounded the wrong way, a terminal value discounted one year too many, a growth rate read as 20 instead of 0.02. In a valuation, those mistakes change the answer — and nobody can tell.
Fairmark splits the job in two:
flowchart LR
D["Financial statements<br/>business plan"]:::input --> L["LLM<br/>extracts the inputs"]:::llm
L --> F["Fairmark<br/>computes · explains · fingerprints"]:::engine
F --> R["LLM<br/>writes the report, citing every step"]:::llm
classDef input fill:#F7F8FA,stroke:#475569,color:#0B1F3A
classDef llm fill:#FFFFFF,stroke:#12325C,color:#0B1F3A
classDef engine fill:#0B1F3A,stroke:#C8A24A,stroke-width:2px,color:#FFFFFF
| Deterministic | Exact decimal arithmetic, fixed solvers. Same input, same output, to the last digit, on every platform. |
| Auditable | Every result carries explain: formula, conventions and each intermediate step — ready for a valuation report. |
| Reproducible | Every result carries an input fingerprint; fairmark verify re-runs it bit for bit. |
| Agent-native | One engine for Python, function calling, MCP, the CLI and Agent Skills for Claude Code and Codex. |
| Built for private companies | Inputs come from financial statements and business plans, not tickers. European practice is first-class. |
| Private by design | The engine never touches the network. Your numbers stay on your machine. |
uv add fairmark # or: pip install fairmark
uv add "fairmark[mcp]" # + MCP serverPrebuilt wheels for Linux (x86_64 and arm64, glibc and musl), macOS (Intel and Apple silicon) and Windows (x86_64): no compiler needed. Requires Python 3.12 or newer.
import fairmark as fm
r = fm.dcf.fcff(
[fm.dcf.given([100, 110, 120])], # one stage: FCFF of years 1 to 3
cash_flow_terms="nominal",
discount_rate_terms="nominal",
wacc="0.10", # as a decimal fraction
timing="end_of_period",
terminal=fm.dcf.gordon("0.02", first_flow="last_flow_grown"),
net_financial_position={"gross_debt": 200, "cash": 0},
)
r.value_of_operations # Decimal('1421.4876033057851239669421487')
r.equity_value # Decimal('1221.4876033057851239669421487')
r.terminal_value_weight # Decimal('0.8086680761099365750528541226')
for s in r.steps: # every step, with the numbers substituted
print(s.label, s.formula, s.value)Period 1: discount factor 1 / (1 + 0.10) 0.9090909090909090909090909091
Period 1: present value 100 × 0.9090909090909090... 90.909090909090909090909090910
Terminal value: Terminal value at N 122.40 / (0.10 − 0.02) 1530
Value of operations 271.9759... + 1149.5116... 1421.4876033057851239669421487
Equity value 1421.4876... − 200 1221.4876033057851239669421487
(excerpt)
Tip. Pass amounts and rates as strings or
Decimalfor exactness. Floats are accepted and converted through their shortest representation (0.1→"0.1").
| Function calling | MCP server |
|---|---|
resp = client.messages.create(
model="claude-opus-5-5",
tools=fm.tools.anthropic(), # or fm.tools.openai()
messages=messages,
)
for block in resp.content:
if block.type == "tool_use":
out = fm.tools.dispatch(block.name, block.input) |
{
"mcpServers": {
"fairmark": {
"command": "uvx",
"args": ["--from", "fairmark[mcp]", "fairmark-mcp"]
}
}
} |
Any provider — Anthropic, OpenAI, Google Gemini, OpenRouter, LiteLLM, and open-weight
models on Ollama, vLLM or LM Studio: fm.tools.for_provider("ollama") returns the right tool
format, fm.tools.dispatch executes the call. LangChain tools and MCP over HTTP (for n8n) are
included. See LLM providers.
Claude Code, Codex and other coding agents — copy skills/fairmark
into your agent's skills folder, or let it drive the CLI:
fairmark methods # what can I compute?
fairmark schema dcf.fcff # which inputs, with units?
fairmark run dcf.fcff --input company.json # JSON in, JSON out
fairmark verify result.json --input company.json # reproduce a result bit for bitNote. Tool definitions are generated from the engine's JSON Schemas, so they never drift from the implementation. Invalid inputs come back as
{"error": {"code": "FM-E004", …}}so the model can correct its own call.
| Method | Status | Description |
|---|---|---|
dates.year_fraction |
Available | Year fraction (day count) |
dates.schedule |
Preview | Schedule of accrual periods |
series.project |
Preview | Value series projected on a timeline |
rates.convert |
Available | Rate conversion between compounding conventions |
rates.real_nominal |
Available | Real and nominal rates (Fisher) |
tvm.npv |
Available | Net present value |
tvm.present_value |
Available | Present value with a constant or time-varying rate |
tvm.xnpv |
Available | Net present value of dated cash flows (XNPV) |
tvm.irr |
Available | Internal rate of return |
tvm.xirr |
Available | Internal rate of return of dated cash flows (XIRR) |
tvm.mirr |
Available | Modified internal rate of return (MIRR) |
tvm.annuity |
Available | Annuity (level or growing, ordinary or due) |
tvm.perpetuity |
Available | Perpetuity (level or growing) |
tvm.future_value |
Available | Future value of a present amount and level payments |
loans.instalment |
Available | Loan instalment (constant instalment, optional balloon) |
loans.number_of_instalments |
Available | Number of instalments of a loan |
loans.amortisation_schedule |
Available | Amortisation schedule (French, constant principal, bullet) |
coc.cost_of_equity |
Available | Cost of equity (CAPM with explicit premia) |
coc.beta_relever |
Available | Beta unlevering and relevering |
coc.beta_peers |
Available | Beta from comparable companies |
coc.cost_of_debt |
Available | Cost of debt |
coc.wacc |
Available | Weighted average cost of capital (WACC) |
tv.gordon |
Available | Terminal value, stable growth (Gordon) |
tv.value_driver |
Available | Terminal value, value driver (growth from reinvestment) |
tv.h_model |
Available | Terminal value, H-model (declining growth) |
tv.exit_multiple |
Available | Terminal value, exit multiple |
tv.limited_horizon |
Available | Terminal value, finite horizon |
tv.normalised_income |
Available | Terminal value, capitalised normalised income |
tv.liquidation |
Available | Terminal value, liquidation |
tv.compare |
Available | Terminal value, comparison of techniques |
dcf.fcff |
Preview | DCF — unlevered free cash flow |
dcf.fcfe |
Preview | DCF — free cash flow to equity |
dcf.free_cash_flows |
Preview | Free cash flow build-up (FCFF and FCFE) |
statements.reorganise |
Preview | Financial statements — operating and financing reorganisation |
ratios.profitability |
Preview | Ratios — profitability |
ratios.capital_structure |
Preview | Ratios — capital structure |
ratios.liquidity |
Preview | Ratios — liquidity |
ratios.working_capital |
Preview | Ratios — working capital |
ratios.debt_service |
Preview | Ratios — debt service |
ratios.covenants |
Preview | Ratios — covenants |
analysis.goal_seek |
Available | Goal seek |
analysis.sensitivity |
Available | Sensitivity table |
analysis.scenarios |
Available | Scenario analysis |
analysis.tornado |
Available | Tornado |
Available: verified independently, with recorded evidence (spec approved, golden cases with sources, numerical audit, code review). Preview: works and is tested; some evidence is still to be recorded. Rule: method status.
What comes next, with its issues and specifications, is in the roadmap, generated from the plan.
flowchart TB
subgraph CORE["fairmark-core · Rust · no I/O"]
direction LR
M["Methods<br/>decimal-exact"]:::core --> E["explain + meta<br/>audit trail · input hash"]:::core
REG["Registry + JSON Schemas<br/>single source of truth"]:::gold
end
REG --> PY["Python SDK<br/><code>pip install fairmark</code>"]:::adapter
REG --> CLI["CLI<br/><code>fairmark run</code>"]:::adapter
PY --> FC["Function calling<br/>Anthropic · OpenAI"]:::agent
PY --> MCP["MCP server<br/>local · stdio"]:::agent
CLI --> SK["Agent Skills<br/>Claude Code · Codex"]:::agent
REG -.-> N8N["n8n node<br/>on the roadmap"]:::later
REG -.-> WASM["WASM playground<br/>on the roadmap"]:::later
classDef core fill:#0B1F3A,stroke:#0B1F3A,color:#FFFFFF
classDef gold fill:#0B1F3A,stroke:#C8A24A,stroke-width:2px,color:#FFFFFF
classDef adapter fill:#12325C,stroke:#12325C,color:#FFFFFF
classDef agent fill:#FFFFFF,stroke:#12325C,color:#0B1F3A
classDef later fill:#F7F8FA,stroke:#94A3B8,color:#475569,stroke-dasharray:4 4
One engine, thin adapters. Every channel is generated from the same registry and schemas: a method added to the core is immediately available to Python, the CLI, MCP and every LLM. Details in docs/en/concepts/architecture.md.
- Decimal arithmetic only — no
f64in any calculation path. - Fixed algorithms; tolerances and iteration counts are reported in the result.
- No silent defaults: unknown or missing inputs are typed errors with stable codes.
- Every result carries
meta.input_hash: same hash + same versions = same result. - Any change to a golden result is a reviewed, versioned
numerical-change.
Golden cases in tests/golden run against both Rust and Python on every
commit. Read the full contract and the project's
constitution.
Fairmark is developed with spec-driven development and is ready for agentic
contributors: every feature starts as a spec with numbered requirements and acceptance
criteria (specs/); every acceptance criterion becomes a test; every PR is
checked by CI for determinism, schema contracts and golden cases, and reviewed by humans and
by AI reviewers with the project's own standards (AGENTS.md).
- Start with CONTRIBUTING.md and issues labelled
good first issue. - Valuation professionals are as welcome as developers: a well-sourced test case is a first-class contribution.
Apache-2.0. The license covers the code; it grants no rights to use the Fairmark name or logo (Apache-2.0, section 6).
Fairmark is a calculation tool, not an advisor: it computes, the user decides. Every decision taken with the help of Fairmark, and every use of its results, belongs exclusively to the user, who chooses the method and the inputs, directly or through an AI agent, and verifies the results before relying on them. Fairmark provides no investment, financial, accounting, tax or legal advice, and is provided without warranty of any kind under the Apache License 2.0. Read the full disclaimer.