Skip to main content

agnara

Capability-native Python for the agentic era.

Agnara is a Python 3.14-native capability framework for services meant to be consumed by humans, applications and AI agents without making HTTP the centre of the architecture. A capability is declared once, with its effects, risk and confirmation requirements, and later exposed through whichever protocol the caller needs.

This distribution is agnara, the capability-first, transport-neutral execution kernel: the capability model, registry, execution context, dependency graph, policies, execution planning and canonical errors. It depends on nothing but the standard library.

Status: 1.0 candidate

The retained A8 publication establishes the historical publication baseline. The project is working toward its first stable public release, 1.0.0; do not infer production readiness, protocol conformance, benchmark leadership or security guarantees without the release evidence. For source migrations, see the repository's docs/MIGRATION_A8_TO_1_0.md.

Which versions are on PyPI is answered by the project page rather than by this file. CHANGELOG.md records what each version contains.

Install

pip install agnara

Pin an actually published version when a build must not move. For the 1.0 candidate, install the synchronized wheels built by the release workflow; do not use A8 as the 1.0 compatibility target.

pip install "agnara==<published-version>"

Requires CPython 3.14 or newer.

Quick start

import asyncio

from agnara import Agnara, Principal, Risk, StandardEffect
from agnara.di import DIContainer, DIRegistry
from agnara.execution import (
    ExecutionContext,
    ExecutionPlan,
    Invocation,
    invoke_result,
)

app = Agnara("billing")


@app.capability(
    description="Refund a captured payment.",
    scopes=("billing:write",),
    effects=(StandardEffect.FINANCIAL_WRITE,),
    risk=Risk.HIGH,
)
def refund(payment_id: str, amount_cents: int) -> str:
    return f"refunded {amount_cents} cents for {payment_id}"


async def main() -> None:
    capabilities = app.compile()
    dependencies = DIRegistry()
    plan = ExecutionPlan.compile(capabilities["billing.refund"], dependencies)

    outcome = await invoke_result(
        plan,
        ExecutionContext(
            Invocation(
                capability_id=plan.definition.id,
                payload={"payment_id": "pay_123", "amount_cents": 2500},
                metadata={},
            ),
            DIContainer(dependencies),
            principal=Principal("quickstart", scopes={"billing:write"}),
        ),
    )
    print(outcome)


asyncio.run(main())

The declared function is returned unchanged, so it stays directly callable and directly testable. Registration is a side effect on the application, not a transformation of the function.

What this release includes

  • capability declaration and a deterministic, freezable registry;
  • stable capability identity, plus effect, risk, idempotency and confirmation metadata;
  • a schema port with a standard-library adapter and compiled per-parameter input validation;
  • dependency injection with compile-time graph validation and scoped resolution;
  • execution plans, direct invocation and optional monotonic deadlines;
  • protocol-neutral policies, principals and scope evaluation;
  • canonical Success / Failure outcomes with stable failure codes;
  • protocol-neutral introspection snapshots and explicit discovery visibility;
  • structured execution telemetry hooks with per-invocation identity.

What it does not include

The HTTP/ASGI, OpenAPI, MCP, CLI and OpenTelemetry functionality lives in separate distributions -- agnara-http, agnara-mcp, agnara-cli and agnara-telemetry -- and is not bundled into this standard-library-only kernel. Each is versioned in step with this one; check its PyPI project page for the versions available to install. Events and A2A remain zero-API reserved namespaces.

Frozen value semantics

Core value types such as CapabilityId and CapabilityDefinition are immutable and slotted. Assigning or deleting either a declared field or an unknown attribute raises dataclasses.FrozenInstanceError; a typo never attaches new state and does not leak CPython's internal slots error.

Confirmation boundary

Capabilities declared with confirmation="required" need an application-provided ConfirmationVerifier when their ExecutionPlan is compiled. Each invocation may carry an explicit opaque ConfirmationEvidence on ExecutionContext; values in generic invocation metadata are not approval.

The verifier receives the exact capability id, invocation, and principal and owns authenticity, input canonicalization, expiry, and replay protection. Missing evidence terminates execution with an interaction request. Rejected evidence terminates it as forbidden. Both outcomes occur before dependency construction or handler effects, and invoke_result() maps them to stable protocol-neutral failure codes.

Telemetry hooks

The core port is agnara.execution.TelemetryHook, with synchronous on_invocation_start(InvocationStartEvent) and on_invocation_terminal(InvocationTerminalEvent) callbacks. Register observers with ExecutionPlan.compile(definition, registry, hooks=[observer]); inheriting from the protocol is optional. Events expose capability, invocation and tracking identity; terminal events also contain monotonic duration and execution outcome, without handler inputs, returned payloads or exception objects.

Both plan construction paths copy the hook collection to a tuple. Missing or non-callable callbacks, coroutine functions and generator functions fail at startup with DefinitionError. Valid callbacks accept one event, return None synchronously and must not block. Their ordinary exceptions are ignored during execution.

Observers own synchronization of their mutable state and must keep their callbacks stable after compilation. Tracking IDs are caller-provided and may repeat; they are not unique span identifiers and should contain no secrets. Exporter startup, flushing and shutdown belong to adapters, not the core runtime. The separate agnara-telemetry package provides metrics and tracing hooks over an application-supplied meter and tracer.

Identity: code constructing InvocationStartEvent or InvocationTerminalEvent supplies invocation_id. Use the same identity for matching start and terminal events; tracking_id is not unique. Hooks that only read events are unaffected, and plans without hooks skip event construction entirely.

License

Apache License 2.0.

Release files for agnara 1.0.2

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

Source distribution (sdist)

Source distribution for agnara 1.0.2
File Size Uploaded
agnara-1.0.2.tar.gz 78.9 kB Details

Built distribution (wheel)

Table of built distributions (wheels) for agnara 1.0.2
File Interpreter ABI Platform
agnara-1.0.2-py3-none-any.whl Python 3 none any Details

Total release size: 185.4 kB

Release files / agnara-1.0.2.tar.gz

Download URL agnara-1.0.2.tar.gz
Size 78.9 kB
Tags Source
SHA-256 checksum
How to use checksums
0256e8d0e4ee61abec8ebd227f760259af753b15f6d6478627c33908d0ce9323
BLAKE2b-256 checksum
How to use checksums
028a62bbf9972dbfb42ff71adbe4f67f37db3a78cf880aeaff60b2b5e6a47786
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 25, 2026.

Transparency log

Release files / agnara-1.0.2-py3-none-any.whl

Download URL agnara-1.0.2-py3-none-any.whl
Size 106.5 kB
Tags Python 3
SHA-256 checksum
How to use checksums
73455aa979f77d94aadba19f3ef9a696169f54a36732008bd34c6f62fb464757
BLAKE2b-256 checksum
How to use checksums
a6c7c08c1571120bc83297399a3e9cf439cbec1f329db7b69a851ed4c099a783
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 25, 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