python/

dotvault — Python bindings

Python bindings for dotvault's public client API. They let a Python program read the per-user secrets a dotvault daemon enrolled and keeps current — talking to the same Vault, resolving a token the same way, and reading from the exact kv/users/<user>/... path dotvault writes to.

The package is a thin ctypes wrapper over dotvault's Go client package compiled to a shared library with go build -buildmode=c-shared. Connectivity, token precedence, the OS-user identity convention, and the path layout all come from the one canonical Go implementation rather than being re-derived in Python, so a consumer can't silently diverge from the daemon. The native bridge imports only the public Go client package — never dotvault's internals.

Scope

This first release exposes the read-only + cached-auth subset of the Go facade:

• Client(config_path=None, identity=None) — load config, build the client.
• authenticate_cached(timeout=None) — resolve a token (DOTVAULT_TOKEN → token file → peer socket borrow when vault.token_socket is configured) and validate it. The socket borrow is a plain read with no browser or prompt, so a host with no local token but a live peer socket authenticates without an interactive login. Never prompts.
• identity_name(), token().
• read_user_secret(service, field, timeout=None), read_kv_field(mount, path, field, timeout=None).

Interactive login (OIDC browser pop, LDAP password + MFA terminal prompts) is deliberately out of scope — driving it across an FFI boundary from inside a Python process is awkward and not what a library caller wants. Provision a token out of band (dotvault login, or the daemon) and these bindings consume it.

Install

Released wheels are published to PyPI (Linux manylinux_2_28, macOS arm64, Windows x86_64):

pip install dotvault          # or: uv pip install dotvault

The wheel is tagged py3-none-<platform>: it carries a native shared library (so it is platform-specific) but contains no CPython C-extension — it is pure ctypes — so a single wheel per OS installs on any Python ≥ 3.9, not one wheel per interpreter version. Its version is derived from the repo's git tags by setuptools-scm (the same tags that version the daemon); dotvault.__version__ reports it.

Wheels are built for glibc Linux x86_64 (manylinux_2_28), Apple-Silicon macOS, and Windows x86_64. There are currently no wheels for Linux aarch64, musl/Alpine, Intel macOS, or Windows arm64 — pip install on those reports "no matching distribution"; build from a checkout instead.

To build from a checkout instead — the tooling is uv:

make python-wheel               # -> python/dist/dotvault-*.whl  (runs `uv build`)
uv pip install python/dist/dotvault-*.whl

or an editable install (requires Go on PATH, since the bridge is compiled on install):

cd python && uv pip install -e .

Building from source needs the Go toolchain and a C compiler — go build -buildmode=c-shared is the one place dotvault uses cgo (CGO_ENABLED=1). The main dotvault binaries remain pure-Go static builds; only this binding links libc. On Windows the C compiler must be a mingw-w64 gcc (the c-shared build does not work with MSVC); GitHub's windows-latest runners ship one, but a local Windows build needs it on PATH.

Concurrency note: set DOTVAULT_TOKEN before the first use of a Client, not concurrently with reads from another thread. The bridge re-reads the facade's env vars from libc on each token-resolving call; that read/write is serialised internally, but it cannot be made safe against the host process mutating its own environment from another thread.

Quick start

import dotvault

# Defaults: dotvault's system config path, identity = the OS user.
with dotvault.Client() as c:
    try:
        c.authenticate_cached(timeout=5)          # env -> token file; no prompt
    except dotvault.LoginRequired:
        raise SystemExit("run `dotvault login` first")
    except dotvault.Unreachable:
        raise SystemExit("vault is unreachable; retry later")

    token = c.read_user_secret("gh", "oauth_token")   # -> str | None
    if token is None:
        raise SystemExit("github enrolment not present")
    use(token)

A read returns the field value, or None when the secret or field is absent — a missing path and a missing field are not distinguished (both are "not there"), matching the Go facade. Transport/authorisation failures raise instead.

Identity is the OS user, not the Vault token

dotvault derives the <user> segment of kv/users/<user>/... from the OS account the process runs as (with any DOMAIN\ prefix stripped), not from the Vault token's display_name or entity. By default a consumer must therefore run as the same OS user as the dotvault that populated the secrets — normally true for a per-user daemon.

If your process runs as a different user (a service account, a container), pass identity= to read that user's secrets:

dotvault.Client(identity="alice")

A wrong identity reads a non-existent path, which surfaces as a None read — not an error.

Errors

Every failure is a DotvaultError or a subclass:

A not-found read is None, never an exception.

Environment variables

The bindings honour the same variables as dotvault: DOTVAULT_TOKEN (a token supplied via the environment) and VAULT_NAMESPACE. VAULT_TOKEN is deliberately ignored — it belongs to the vault CLI and must not leak into dotvault's session; use DOTVAULT_TOKEN.

The Go runtime snapshots its environment at load time, so the bridge re-reads these two variables from the live process environment on each call. That means setting os.environ["DOTVAULT_TOKEN"] after import dotvault works as you'd expect.

Development

make python-lib     # build the native bridge into the package for local use
make python-test    # build the bridge + run the Python test suite (via `uv run`)
go test ./python/bridge/...   # the Go-side bridge unit tests

make python-test and make python-wheel shell out to uv, which provisions Python and the test/build dependencies; only the Go toolchain and a C compiler need to be present beforehand.

The Python tests run fully offline — they point at a closed Vault port and assert the error categorisation — so no live Vault is needed.

Releasing

Publishing a GitHub Release with a vX.Y.Z tag drives both the Go release (release.yml) and the Python wheels: the .github/workflows/python.yml publish job builds the per-OS wheels and uploads them to PyPI via the official pypa/gh-action-pypi-publish action using Trusted Publishing (OIDC — no stored token). One-time setup before the first release: register a pending publisher on PyPI for project dotvault pointing at this repository, workflow python.yml, and environment pypi. A dev/untagged build produces a PEP 440 local version that PyPI rejects, so only exact tagged releases publish. Consider a one-off TestPyPI dry run (temporarily pointing the action at TestPyPI) before the first real release to validate the OIDC/publisher wiring.