Skip to main content

macula-py

CI License Python GitHub Sponsors

Macula

A Python node on the macula 12 mesh, over macula-go's C ABI


Status, 2026-09-26: on the macula 12 wire: ML-DSA-87 identities (as the ML-DSA-87 + RSA-PSS-4096 composite in pq_hybrid, the fleet's profile), ML-KEM hybrid key exchange, signed requests. Calls and streams by direct dial, serving (under an org or in a node's own namespace, open or gated on a post-quantum UCAN), publish/subscribe, the DHT and node-served content are tested against two in-process macula 12 stations on every CI run. Device request proofs (realm join) and ownership proofs are held to their verifiers' vectors, and checked once through the realm's and mcl_om's own verifiers. 0.1.0 spoke the retired classical wire and cannot reach the current fleet.

What is this?

A Python SDK for the Macula mesh: a node's key, a pool of links to stations it pins by node_id, calls and streams that reach a provider by direct dial, serving procedures, publish/subscribe, the DHT and node-served content, all as asyncio coroutines.

Macula is a federated mesh for sovereign application networks. A station relays and holds the DHT; a node is anything else that joins, and this package is a node.

It is a binding, not a reimplementation. The protocol lives in macula-go, which exports it as a C ABI (cabi/macula.h, contract in cabi/CONTRACT.md) shared by every SDK not written in Go. macula-py loads that library with ctypes, so the wire, the post-quantum handshake and the signing rules are the ones macula-go already checks byte for byte against macula itself.

Install

pip install macula-py

Wheels exist for Linux x86-64 and arm64 (glibc 2.28 or later), macOS 13 or later on Apple silicon and Intel, and Windows x86-64. Each carries macula-go's library; nothing compiles at install time and there are no runtime Python dependencies. On any other platform pip finds no wheel.

Quick start

import asyncio

from macula_py import NodeKey, Pool, Seed

STATION = Seed("station-fi-helsinki.macula.io", 4433,
               "004d1f470097ccf8826ce291900e882fdb1f20375e53901facaec0f23eb4efd8")
REALM = "abb81b5a614b63551b400b810648c0c8a78efad845442630c94b46cc95d2fcd1"  # io.macula
REALM_KEY = "..."  # io.macula's public realm key, hex


async def main() -> None:
    key = await NodeKey.load_or_create("node.key")  # pq_hybrid; the puzzle takes a second
    async with await Pool.connect(key, [STATION], realm_trust={REALM: REALM_KEY}) as pool:
        print(await pool.call(REALM, "mcl-echo/echo", "hello"))


asyncio.run(main())

A seed is pinned: the station must prove the node_id you give. A realm's key decides which advertisements in it you trust; procedures in a node's own namespace (~<node_id>/<name>, see Pool.own_procedure) need none.

The API

NodeKey generate, load, load_or_create, save, node_id, public_key, profile, sign, verify, free; ucan, device_request_proof, ownership_proof (below)
Pool.connect seeds, realm_trust, and the pool's tuning; async with closes it
calls call, providers; call(..., ucan=, proofs=) presents a UCAN
serving serve(realm, procedure, handler, policy=None): handler(request) returns the result, directly or as an awaitable; an exception reaches the caller as a ProviderError of code handler_error
streams open_stream, serve_stream; a Stream has send, send_value, close_send, reply, abort, close, recv, and iterates its frames
pub/sub publish, subscribe; a Subscription iterates its events, or next(timeout_ms)
content share_content, unshare_content, get_content
DHT find_record, find_records, find_records_by_type, put_record

UCANs

A procedure served with a policy answers only callers presenting a UCAN (macula 12's post-quantum capability token) the policy accepts; the provider checks each call and stream open before the handler sees it, as macula does, and answers the rest unauthorized (or malformed_frame for a proof no token in the chain names):

from macula_py import UcanRequired
from macula_py.ucan import proof_id

served = await provider.serve(realm, procedure, handler, policy=UcanRequired(root.node_id()))

token = root.ucan(caller.node_id(), [{"with": "mri:org:io.macula/acme", "can": "invoke"}],
                  exp=int(time.time()) + 3600)
await caller.call(realm, procedure, payload, ucan=token)

# Delegated: alice hands the caller one procedure, naming her grant as its parent.
sub = alice.ucan(caller.node_id(), [{"with": "mri:proc:io.macula/acme/count_v1", "can": "invoke"}],
                 exp=int(time.time()) + 600, prf=[proof_id(to_alice)])
await caller.call(realm, procedure, payload, ucan=sub, proofs=[to_alice])

A token is minted for the node that will present it. RealmMemberRequired(key_id, can) gates on a realm key instead, named by macula_py.ucan.key_id(realm_public_key, profile) (the key as carried, bytes). macula's test/vectors/UCAN_V1.md is the contract.

Proofs for a realm and for a service

NodeKey.device_request_proof(realm, procedure, request, rule) signs a device's request to a realm (realm proof v2): a join session's body (rule="http", a mapping or the body text exactly as sent) or a membership UCAN request over the mesh (rule="mesh"). NodeKey.ownership_proof(realm, procedure, payload) returns the payload with the asserted_by block that authorises its fields to a service such as mcl_om; send it as the payload. Neither signs a "caller": the caller is the verified signer, so a request or payload carrying one is refused.

Payloads are what macula's wire carries: str, int within int64, float, None, bytes, lists and dicts with str keys. There is no boolean on the wire: send 1 and 0. A Python bool is refused before it leaves, since Python treats it as an int.

Every networked method is a coroutine, and each native call runs on a thread of its own, so long waits never starve other calls. The methods that take timeout_ms, and every wait for an event, a served call or a stream frame, carry a cancel token: cancelling the task (or asyncio.wait_for timing out) ends the native call at once. The rest (publish, subscribe, serve, stop, close, and a stream's sends and ends) are short native calls without one; cancelling stops the waiting and the call runs to its end.

Errors are typed: ProviderError, RelayError, StreamError, NotSharedError, ContentUnavailableError, NoProviderError, MaculaTimeoutError (also a TimeoutError), InvalidArgumentError (also a ValueError), ClosedError, RefusedError, all MaculaError.

Development

Needs Go 1.27 and a C compiler to build macula-go's library locally.

python -m venv .venv && .venv/bin/pip install -e ".[dev]"
eval "$(scripts/build_native.sh)"   # the library and teststation, from abi/MACULA_GO_REF
.venv/bin/pytest

abi/macula.h is macula-go's header at the ref in abi/MACULA_GO_REF; tests/test_abi_declarations.py holds the ctypes declarations to it function for function, and build_native.sh refuses a header that differs from the ref's.

scripts/live_check.sh runs tests/live/ against one fleet station (helsinki and io.macula by default) with keys made for the run and never saved. It publishes once, and advertises one UCAN-gated procedure in its own namespace under a throwaway realm. CI never runs it.

scripts/interop/ownership_proof.sh and scripts/interop/device_request.sh check proofs this binding signs against the verifiers themselves: mcl_om's (in macula's pinned CI image), after the payload has crossed a station as a provider receives it, and the realm's. See scripts/interop/README.md.

A v* tag publishes to PyPI through Trusted Publishing, using only macula-go's released libraries, each checked against the release's SHA256SUMS and its build provenance attestation.

License

Apache-2.0. See LICENSE.


Built on macula-go, for Python -- sponsor the work if this saved you some time

Metadata

Release files for macula-py 0.3.1

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Built distributions (wheels)

Table of built distributions (wheels) for macula-py 0.3.1
File
macula_py-0.3.1-py3-none-win_amd64.whl Python 3 none Windows x86-64 Details
macula_py-0.3.1-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl Python 3 none Linux glibc 2.17+ ARM64, Linux glibc 2.28+ ARM64 Details
macula_py-0.3.1-py3-none-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl Python 3 none Linux glibc 2.5+ x86-64, Linux glibc 2.28+ x86-64 Details
macula_py-0.3.1-py3-none-macosx_13_0_x86_64.whl Python 3 none macOS 13.0+ x86-64 Details
macula_py-0.3.1-py3-none-macosx_13_0_arm64.whl Python 3 none macOS 13.0+ ARM64 Details

Total release size: 17.0 MB

Release files / macula_py-0.3.1-py3-none-win_amd64.whl

Download URL macula_py-0.3.1-py3-none-win_amd64.whl
Size 3.5 MB
Tags Python 3 Windows x86-64
SHA-256 checksum
How to use checksums
d82287332c605235e3671cd4b120429c8f05fdc842eb41ca7114159359253455
BLAKE2b-256 checksum
How to use checksums
56cd184246d6dc6392ffd7427cb6972c932bac9e49fe04fdd904f6bd94f267e7
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 28, 2026.

Transparency log

Release files / macula_py-0.3.1-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl

Download URL macula_py-0.3.1-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl
Size 3.2 MB
Tags Linux glibc 2.17+ ARM64 Linux glibc 2.28+ ARM64 Python 3
SHA-256 checksum
How to use checksums
8bff6aca55147b8daa2827c12c790d36f4c94f24dd446a3572f10ccfd1d3af48
BLAKE2b-256 checksum
How to use checksums
43d328725bd0d0c714d856e2f2a24d10b6ff36d6e441de98a772237828089790
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 28, 2026.

Transparency log

Release files / macula_py-0.3.1-py3-none-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl

Download URL macula_py-0.3.1-py3-none-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl
Size 3.6 MB
Tags Linux glibc 2.28+ x86-64 Linux glibc 2.5+ x86-64 Python 3
SHA-256 checksum
How to use checksums
289ffc876f08d42fc23f482313a64854b4d93018b13876624674085d107579a3
BLAKE2b-256 checksum
How to use checksums
cdb7b588e228b1160332d16ffb9e42bda030dbb14ca0a7a8e2626e134c91e2c7
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 28, 2026.

Transparency log

Release files / macula_py-0.3.1-py3-none-macosx_13_0_x86_64.whl

Download URL macula_py-0.3.1-py3-none-macosx_13_0_x86_64.whl
Size 3.5 MB
Tags Python 3 macOS 13.0+ x86-64
SHA-256 checksum
How to use checksums
068f73b6670af163a1e1229572777a9ebed75c280ab10204cf41309822db64b1
BLAKE2b-256 checksum
How to use checksums
6c494eec185d3e02cb4f15b6ccbc475446a700819d48d6bb474177317586975c
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 28, 2026.

Transparency log

Release files / macula_py-0.3.1-py3-none-macosx_13_0_arm64.whl

Download URL macula_py-0.3.1-py3-none-macosx_13_0_arm64.whl
Size 3.2 MB
Tags Python 3 macOS 13.0+ ARM64
SHA-256 checksum
How to use checksums
ebbda735c7ac78fdc877c92bddfa32dc16c71bd31dd24e53eb53da3a771ae1e6
BLAKE2b-256 checksum
How to use checksums
d1347fb9812b199a5a9db975200fffc141755e20406f6431a6c6e7d1e36a2088
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 28, 2026.

Transparency log
Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page