fmp-py-sdk

fmp-py-sdk is the Python distribution for the Rust-backed fmp package, a typed client for the Financial Modeling Prep (FMP) data API.

The client exposes every libfmp endpoint method: 271 methods grouped into 30 domain namespaces on FmpClient. Each namespace is generated from a registry that is validated against the real libfmp signatures, so the Python surface cannot drift from the Rust one.

python -m pip install fmp-py-sdk

from fmp import FmpClient

client = FmpClient()
rows = client.quote.short("AAPL")
print(rows[0].symbol, rows[0].price)

FmpClient() reads the FMP_API_KEY environment variable when token is omitted (unset, empty, or whitespace-only counts as absent); an explicit token=... always wins, and auth_mode="none" ignores the variable. With neither a token nor the variable, the default host raises FmpConfigError naming FMP_API_KEY, while a custom base_url selects no auth.

FmpClient is synchronous: each call releases the Python GIL while the async Rust transport waits, so threads keep running. A Python async facade is not part of this release.

Requirements

• CPython 3.10 or newer (abi3-py310: one wheel per platform covers every supported interpreter).
• Rust is only needed to build from source; a built wheel is self-contained.

Namespaces

Endpoints live under one attribute per domain, and nested domains such as statements group their sub-namespaces. Required arguments are positional, optional ones are keyword-only, and every name is snake_case: the FMP wire casing (sicCode, linkXlsx) never reaches Python.

from fmp import FmpClient

client = FmpClient()

quotes = client.quote.short("AAPL")
income = client.statements.income.statement("AAPL", period="annual", limit=5)
valuation = client.dcf.custom_discounted_cash_flow(
    "AAPL", beta=1.2, tax_rate=0.21, long_term_growth_rate=4.0
)

statements nests as_reported, balance, cash_flow, growth, income, metrics, ratios, reports, segmentation, and summaries.

Responses

• Typed rows: most methods return list[Model], where each model is a generated, immutable, picklable class living in the domain package (for example fmp.quote.QuoteShort or fmp.statements.income.IncomeStatement). Dates and naive timestamps are datetime.date and datetime.datetime; RFC 3339 timestamps (for example TipRanksRatingSearchResult.date) stay str with the exact wire text. Shared Rust response types map to one shared Python class rather than a copy per endpoint.
• Dynamic rows: endpoints whose documented shape is open-ended return list[dict[str, Any]] with the raw provider keys, for example client.sec_filings.search_industry_classifications(symbol="AAPL"). A dynamic field inside a typed model (such as FinancialReportJson.sections) is exposed as Any.
• Binary bodies: client.statements.reports.xlsx("AAPL", 2022, "FY") returns a single fmp.BinaryPayload instead of a list. data is the body as bytes, alongside content_type, content_disposition, and byte_len.

Validation and errors

Arguments are validated locally before any request: an empty ticker, a malformed date, or a non-finite float raises FmpValidationError whose message starts with the argument name, and nothing is sent. The exception hierarchy lives in fmp.errors (every class is also exported from the package root):

from fmp import FmpClient
from fmp.errors import FmpStatusError

client = FmpClient()
try:
    client.quote.short("AAPL")
except FmpStatusError as error:
    print(error.status, error.endpoint, error.body)

Bodies attached to errors are redacted before they reach Python: an echoed apikey query value shows as [REDACTED].

Custom router or proxy

import os

from fmp import FmpClient

client = FmpClient(
    base_url=os.environ["FMP_PROXY_BASE_URL"],
    path_prefix="router/stable",
    auth_mode="custom_header",
    auth_name="X-Proxy-Token",
    auth_prefix="Bearer ",
    token=os.environ["FMP_PROXY_TOKEN"],
    headers={"X-Tenant": os.environ["FMP_TENANT"]},
)
rows = client.quote.short("AAPL")

Available auth modes are none, fmp_header, fmp_query, bearer, custom_header, and custom_query. auth_mode="none" supports credential-free local or trusted routers. Redirect following is either disabled or same-origin only.

Typing

The package ships a py.typed marker, and every native module has a .pyi stub generated by pyo3-stub-gen from the Rust signatures, so pyright and mypy see the exact method names, keyword-only arguments, and return types. The public modules follow a conventional HTTP SDK layout: FmpClient lives in fmp.client, the exception hierarchy in fmp.errors, and the models and namespace classes in the domain packages such as fmp.quote and fmp.statements.income. Transport, configuration, and runtime internals are intentionally not exposed as Python modules.

Secret URLs

FinancialReportDate.link_json / link_xlsx (the fmp.statements.reports model listing available financial reports) are download links that embed your API key. The Python model treats them the way FmpClient treats danger_allow_insecure_authentication: the unsafe path exists, but it is conspicuous.

• The links are not attributes. Read one with row.expose_secret_url_json() or row.expose_secret_url_xlsx() and treat the value as a credential.
• repr(row) and str(row) print link_json=[REDACTED URL], so a stray print or log line never leaks the key.
• For local debugging only, fmp.set_reveal_secret_urls(True) reveals the URLs in repr() process-wide (fmp.reveal_secret_urls() reads the flag back). Setting FMP_REVEAL_SECRET_URLS=1 (also true / yes, case-insensitive) before the first use turns it on at start-up.
• Pickles of the model contain the real URLs: pickle.dumps must round-trip the value, so store them as carefully as the key itself.

Regenerating the response models

The pyo3 response models under crates/fmp-py/src/models/ are generated from the libfmp response structs by the gen_models binary in the sibling crates/fmp-py-gen crate. Run it from the repository root:

cargo run -p fmp-py-gen --bin gen_models

The generator only depends on syn, not on fmp-py, so it still builds and runs when fmp-py fails to compile against a changed libfmp. Regeneration must be idempotent: git status crates/fmp-py/src/models is clean after a second run.

Validating the endpoint registry

The Python namespaces are described by one TOML file per domain under crates/fmp-py-gen/registry/ (quote.toml for client.quote). Each entry names the Python method, the libfmp Client method, its query type, the constructor arguments and optional with_* setters with their arg kinds, and the generated response model. Validate the registry against the real libfmp signatures and the generated models from the repository root:

cargo run -p fmp-py-gen --bin registry_check

The check parses crates/libfmp/src/endpoints/** with syn, so an unknown method, a mismatched query type, an unknown arg kind, a setter that does not exist, or a missing model file fails with the file and entry named. Query types emitted by macro_rules! are recovered by expanding the macro; the report says whether each entry was verified directly, through a macro, or trusted because its constructor could not be seen, and ends with the total (registry ok: 271 verified, 0 trusted). Pass a directory argument to check a different registry tree.

Regenerating the endpoint namespaces

The namespace classes under crates/fmp-py/src/namespaces/ (QuoteNamespace for client.quote) are generated from the same registry by the gen_namespaces binary, which validates the registry first and emits nothing on any error. Run it from the repository root:

cargo run -p fmp-py-gen --bin gen_namespaces

Flat domains become namespaces/<domain>.rs; a nested path such as statements.income becomes namespaces/statements/mod.rs (the parent with a getter per sub-namespace) plus namespaces/statements/income.rs. Required arguments are positional, optional ones keyword-only, and an argument named after a Python keyword (from) is spelled with a trailing underscore (from_) on the Python side while the libfmp setter keeps its name. Entries marked binary = true (the endpoints whose libfmp method returns BinaryResponse, such as the XLSX financial report download) return a single BinaryPayload instead of a list of models; entries marked response = "dynamic" return list[dict[str, Any]]. BinaryPayload is hand-written in src/binary.rs and exported from the package root as fmp.BinaryPayload.

The same run emits the wiring that binds the generated code into the extension, so adding a domain never edits lib.rs or client.rs:

The hand-written src/facade.rs keeps only the top-level fmp surface.

Regenerating the stubs and public packages

The .pyi stubs under python/fmp/_native/ and the public python/fmp/<path>/__init__.py packages are written by pyo3-stub-gen through the stub_gen binary, from the repository root:

cargo run -p fmp-py --bin stub_gen

stub_gen post-processes each __init__.pyi (absolute imports, no # ruff: noqa header, ruff format --isolated --line-length 88 in .py mode) so a run on an unchanged tree leaves git status clean. It needs ruff at the version pinned in this crate's .pre-commit-config.yaml and as RUFF_VERSION in src/bin/stub_gen.rs (currently 0.15.12, the two must agree): either that exact ruff on PATH or uvx, which fetches it. Run it after gen_models or gen_namespaces, and commit the result.

This project is available under the MIT License.