Python SDK
Resolve SecretSpec secrets from Python
The Python SDK (secretspec) is a thin client over a pyo3 extension that calls secretspec::resolve_json directly. Resolution (providers, chains, profiles, generation, as_path) happens in the Rust core, so the SDK inherits every provider with no Python-side logic.
Quick start
from secretspec import SecretSpec
resolved = (
SecretSpec.builder()
.with_provider("keyring://")
.with_profile("production")
.with_reason("boot web app")
.load()
)
print(resolved.provider, resolved.profile)
db = resolved.secrets["DATABASE_URL"]
print(db.get) # the value, or the file path for as_path secrets
resolved.set_as_env() # export everything into os.environ
A missing required secret raises MissingRequiredError; any other failure raises SecretSpecError (with a stable .kind).
Caller context
New in version 0.20
from secretspec import CallerContext, SecretSpec
builder = SecretSpec.builder().with_caller(CallerContext(
name="git",
version="2.51.0",
operation="credential_get",
resource="github.com",
))
Caller context identifies the invoking integration in audit records but never satisfies require_reason. Do not put credentials or secret values in it.
Inline specifications
New in version 0.20
Use .with_inline_spec(spec, base_dir) to resolve a strict inline-spec v2 dictionary (SecretSpec 0.21+) without a manifest file. base_dir resolves relative provider paths. An inline call uses the versioned native entry point, so it cannot fall back to a filesystem manifest on an older runtime.
Scopes
New in version 0.17
Use .with_scope("api") to resolve only a named [scopes.api] subset. The selected name is available as resolved.scope and report.scope:
resolved = SecretSpec.builder().with_scope("api").load()Typed access (codegen)
Generate typed classes with secretspec schema plus quicktype, then build them from resolved.fields():
$ secretspec schema | quicktype -s schema --top-level SecretSpec --lang python -o secrets_gen.pyfrom secrets_gen import SecretSpec as Secrets # typed
typed = Secrets.from_dict(resolved.fields())
print(typed.database_url) # typed strNative library
The resolver is statically linked into a pyo3 extension (secretspec._native, built from the secretspec-py-native crate) using pyo3’s abi3-py39 feature, so the published cp39-abi3 wheel is self-contained — there is no separate cdylib to locate and no runtime dlopen.
Last updated Oct 11, 2026