Cedar Python

cedarpy helps you use the (Rust) Cedar Policy library from Python. You can use cedarpy to:

• check whether a request is authorized by the Cedar Policy engine
• validate policies against a schema
• format policies

cedarpy releases correspond to the following Cedar Policy engine versions:

Beginning with v4.1.0, cedarpy's version number indicates the Cedar Policy engine major and minor version that it is based on. cedarpy increases the patch number when releasing backwards-compatible changes and bug fixes. So the cedarpy and Cedar Engine patch versions can and will diverge. Select the cedarpy version that provides the Cedar Policy language and engine features you need.

cedarpy packages are available for the following platforms:

Note: This project is not officially supported by AWS or the Cedar Policy team.

Using the library

Releases of cedarpy are available on PyPi. You can install the latest release with:

pip install cedarpy

(See the Developing section for how to use artifacts you've built locally.)

Authorizing access with Cedar policies in Python

Now you can use the library to authorize access with Cedar from your Python project using the is_authorized function. Here's an example of basic use:

from cedarpy import is_authorized, AuthzResult, Decision

policies: str = "//a string containing cedar policies"
entities: list = [  # a list of Cedar entities; can also be a json-formatted string of Cedar entities
    {"uid": {"__entity": { "type" : "User", "id" : "alice" }}, "attrs": {}, "parents": []}
    # ...
]
request = {
    "principal": 'User::"bob"',
    "action": 'Action::"view"',
    "resource": 'Photo::"1234-abcd"',
    "context": {}
}

authz_result: AuthzResult = is_authorized(request, policies, entities)

# so you can assert on the decision like:
assert Decision.Allow == authz_result.decision

# or use the 'allowed' convenience method 
assert authz_result.allowed

# or even via AuthzResult's attribute subscripting support 
assert authz_result['allowed']

The AuthzResult class also provides diagnostics and metrics for the access evaluation request.

See the unit tests for more examples of use and expected behavior.

Authorize a batch of requests

You can also authorize a batch of requests with the is_authorized_batch function. is_authorized_batch accepts a list of requests to evaluate against shared policies, entities, and schema.

Batch authorization is often much more efficient (+10x) than processing authorization requests one by one with is_authorized. This is because the most expensive part of the authorization process is transforming the policies, entities, and schema into objects that Cedar can evaluate. See RFC: support batch authorization requests for details.

Here's an example of how to use is_authorized_batch and the optional request-result correlation_id:

batch_id:str = randomstr()
requests: List[dict] = []
for action_name in action_names:
    requests.append({
        "principal": f'User::"{user_id}"',
        "action": f'Action::"{action_name}"',
        "resource": f'Resource::"{resource_id}"',
        "context": context_keys,
        "correlation_id": f"authz_req::{batch_id}-{action_name}"
    })

# ... resolve get policies, entities, schema ...

# process authorizations in batch
authz_results: List[AuthzResult] = is_authorized_batch(requests=requests, policies=policies, entities=entities, schema=schema)

# ... verify results came back in correct order via correlation_id ...
for request, result, in zip(requests, authz_results):
    assert request.get('correlation_id') == result.correlation_id

cedar-py returns the list of AuthzResult objects in the same order as the list of requests provided in the batch.

The above example also supplies an optional correlation_id in the request so that you can verify results are returned in the correct order or otherwise map a request to a result.

Reusing parsed policies for performance

Parsing the policy set (PolicySet::from_str) is the dominant per-call cost in is_authorized. When your policies are static, for example a code-checked policy set loaded once at startup in a long-running service or AWS Lambda, you can parse them a single time into a reusable PolicySet handle and pass that handle wherever you'd pass a policies string. This skips the re-parse on every call.

from cedarpy import PolicySet, is_authorized, is_authorized_batch, Decision

policies: str = "// a string containing cedar policies"

# Parse once (e.g. at process/Lambda cold start). Parse errors raise ValueError here,
# rather than being folded into an authorization result.
policy_set = PolicySet.from_str(policies)

# Reuse the handle across many requests — no re-parse per call:
authz_result = is_authorized(request, policy_set, entities)
authz_results = is_authorized_batch(requests, policy_set, entities)

The PolicySet handle is accepted anywhere a policies string is accepted: is_authorized, is_authorized_batch, and is_authorized_partial. Passing a plain string still works exactly as before; the handle is purely opt-in and fully backwards compatible. The handle's memory is released automatically when the last Python reference is dropped.

A PolicySet can also be built from the Cedar JSON (EST) policy format with PolicySet.from_json_str(...), and supports len(policy_set) (number of policies) and str(policy_set) (the policy set rendered back to Cedar text).

This complements batch authorization: batching amortizes entity/schema parsing across a set of requests evaluated together, while a reusable PolicySet amortizes policy parsing across calls made at different times. The two compose. The speedup grows with policy size — in the project's benchmark suite (release build, ~10-entity calls) reuse is ~1.3x on a single-rule policy but ~9x on a typical production-scale policy (~16 KB / 60 rules), where it removes ~1.4 ms of policy parsing per call.

On a successful evaluation the result's metrics reflect the reuse: parse_policies_duration_micros measures only the (near-zero) borrow rather than the original parse, and metrics["policies_pre_parsed"] is 1 (it is 0 when policies are passed as a string and parsed on that call). As with all metrics, these keys are present only on successful evaluations — an error result carries an empty metrics map.

Advanced: adding per-request policies to a static base. Most callers reuse a single static PolicySet as above. If, however, your policy set is mostly static but a few policies vary per request, you don't have to re-parse the whole base text each time — PolicySet.with_added_str(fragment) clones the compiled base and parses only the fragment, returning a new handle (the base is left unchanged):

base = PolicySet.from_str(static_policies)        # parse the large static base once

# per request: add only the small dynamic fragment — the base is not re-parsed
for_request = base.with_added_str(per_request_policies)
authz_result = is_authorized(request, for_request, entities)

The result is equivalent to authorizing against the concatenated base-plus-fragment text. (Cedar assigns surface-syntax policies a positional PolicyId — policy0, policy1, … — per parse, so a fragment parsed on its own restarts at policy0; with_added_str renumbers the fragment's colliding ids to follow the base, exactly as concatenation would, and any @id("...") annotations are preserved and still resolve via diagnostics.id_annotations_by_reason.)

Reusing parsed entities for performance

Entities are parsed on each call too: is_authorized deserializes the entities JSON and computes the transitive closure of the parents graph. When you authorize many requests against a large, stable entity graph, you can parse it once into a reusable Entities handle and pass it wherever you'd pass an entities string or list:

from cedarpy import Entities, PolicySet, is_authorized

base = Entities.from_json_str(entities_json)       # parse the stable graph once
authz_result = is_authorized(request, policy_set, base)

The common shape is a large, stable base graph (your organization's users and groups, say) plus a small set of entities that are specific to each request (the resources being acted on). Build the base once and merge the per-request delta with with_added_json_str, which parses only the delta and returns a new handle — the base is immutable and reused across every request:

base = Entities.from_json_str(org_graph_json)      # users, groups: parsed once, reused

for request in requests:
    # add only this request's entities (e.g. the documents involved); base is not re-parsed
    for_request = base.with_added_json_str(request_entities_json)
    is_authorized(request, policy_set, for_request)

with_added_json_str is a disjoint union: a delta entity whose uid already exists in the base (and is not identical) raises ValueError. An optional schema argument to from_json_str / with_added_json_str validates the entities at construction. The handle is accepted anywhere an entities string/list is accepted (is_authorized, is_authorized_batch, is_authorized_partial), supports len() and str(), and sets metrics["entities_pre_parsed"] to 1 on the reuse path. As with the PolicySet handle, passing a string or list still works unchanged — the handle is purely opt-in.

Note: the Entities, PolicySet, and Schema handles are independent and compose — reuse whichever inputs are stable for your workload, or all three.

Reusing parsed schema for performance

The schema is parsed on each call too. When your schema is stable (it usually is), you can parse it once into a reusable Schema handle and pass it wherever you'd pass a schema string:

from cedarpy import Schema, PolicySet, Entities, is_authorized

schema = Schema.from_str(cedar_schema_text)          # Cedar human-readable syntax
# or
schema = Schema.from_json_str(cedar_schema_json)     # Cedar JSON format

authz_result = is_authorized(request, policy_set, entities, schema=schema)

The handle is accepted anywhere a schema string or dict was accepted: is_authorized, is_authorized_batch, is_authorized_partial, validate_policies, and the schema argument to Entities.from_json_str / Entities.with_added_json_str. Passing a plain string or dict still works unchanged — the handle is purely opt-in.

When a Schema handle is used, the authorization response sets metrics["schema_pre_parsed"] to 1 (vs. 0 when a string is parsed on the fly).

Linking policy templates

A Cedar policy template is a policy with ?principal / ?resource slots. A slot is a placeholder that marks where a principal or resource is filled in later, when the template is linked to concrete entities to produce a real, evaluatable policy. So a template is a rule written once and linked per use. The canonical case is per-principal and per-resource grants — allow this person to view this photo while their subscription is active. You write that rule once as a template, then link it per grant.

PolicySet.from_str already parses templates; they just authorize nothing until linked. with_linked fills a template's slots and returns a new PolicySet handle (immutable, like with_added_str — the base is left unchanged):

from cedarpy import PolicySet, is_authorized, Decision

# two grant rules: an active subscriber may view a granted photo, and
# editing also requires a non-free plan (trial users included)
base = PolicySet.from_str("""
    @id("photo-access")
    permit (
      principal == ?principal,
      action == Action::"view",
      resource == ?resource
    )
    when { principal.subscriptionActive };

    @id("photo-edit")
    permit (
      principal == ?principal,
      action == Action::"edit",
      resource == ?resource
    )
    when
    { principal.subscriptionActive && (!principal.freePlan || principal.inTrial) };
""")

# fill the slots to grant alice access to one photo
linked = base.with_linked(
    template_id="photo-access",          # which template to fill in
    new_id="alice-vacation",             # id for this link (the filled slots, which act as a policy)
    values={"?principal": 'User::"alice"', "?resource": 'Photo::"vacation.jpg"'},
)

# alice's subscription is active, so she may view the photo
entities = [{"uid": {"type": "User", "id": "alice"},
             "attrs": {"subscriptionActive": True}, "parents": []}]
request = {"principal": 'User::"alice"', "action": 'Action::"view"', "resource": 'Photo::"vacation.jpg"'}

result = is_authorized(request, linked, entities)
assert result.decision == Decision.Allow

# the decision points at the link's own id; the template's @id comes alongside it
result.diagnostics.reasons                    # ['alice-vacation']
result.diagnostics.id_annotations_by_reason   # {'alice-vacation': 'photo-access'}

Linking produces a template-linked policy — the slots filled with concrete values, which evaluates as a policy in its own right — and new_id is the id assigned to it. Linking does not rename or consume the template: the template stays in the set, so you can link it again under a different new_id to grant another principal (that's the whole point — one template, many grants). new_id is yours to choose; it is neither a template id nor a principal. (This mirrors the Cedar CLI's link --template-id … --new-id ….)

To create many template-linked policies at once, use with_linked_batch. This is the primary linking path: in a single call it fills the slots of any templates in the PolicySet — the same one repeatedly or several different ones — with the entities you choose. If any link in the batch fails, the whole call raises and no handle is returned. It clones the base once rather than per link:

linked = base.with_linked_batch([
    # the same template, linked twice — two view grants
    {
        "template_id": "photo-access",
        "new_id": "alice-vacation",
        "values": {
            "?principal": 'User::"alice"',
            "?resource": 'Photo::"vacation.jpg"',
        },
    },
    {
        "template_id": "photo-access",
        "new_id": "bob-skyline",
        "values": {
            "?principal": 'User::"bob"',
            "?resource": 'Photo::"skyline.jpg"',
        },
    },
    # a different template, but using dict assignment — an edit grant
    {
        "template_id": "photo-edit",
        "new_id": "alice-edit-vacation",
        "values": {
            "?principal": {"type": "User", "id": "alice"},
            "?resource": {"type": "Photo", "id": "vacation.jpg"},
        },
    },
])

A slot value is given as a Cedar string ('User::"alice"') or a {"type": ..., "id": ...} dict — the same two forms is_authorized accepts for a request principal/resource. A linked PolicySet is still just a PolicySet, so it works everywhere a policies string or handle does (is_authorized, is_authorized_batch, is_authorized_partial), and existing callers are untouched. Note that the slots fill only the principal and resource; the when { principal.subscriptionActive } condition is fixed in the template and shared by every link — write the rule (and its conditions) once, vary only who and what.

Identify a template by its @id, not its positional id. Cedar assigns a policy id by position — policy0, policy1, … — per parse. That position can shift: with_added_str, for example, renumbers an incoming fragment's ids when it layers them onto a base. A template's @id annotation is invariant across parsing and merging, so it is the stable key to link against. with_linked accepts either — it matches the literal id first, then the @id — but prefer the @id. The match must be unambiguous; two templates sharing one @id raises rather than guessing. (An @id is otherwise inert and is not the template's id; cedarpy resolves it the way the Cedar CLI's link command does.)

templates() returns the full picture — each template's ids, the slots it needs, and the links (fillings) produced from it:

linked.templates()
# [{'id': 'policy0',                       # positional id (the fragile one)
#   'id_annotation': 'photo-access',       # the @id — the stable key to link by
#   'slots': ['?principal', '?resource'],  # the slot keys a link must fill
#   'links': [                             # the concrete policies produced from this template
#       {'id': 'alice-vacation',
#        'values': {'?principal': 'User::"alice"', '?resource': 'Photo::"vacation.jpg"'}},
#       {'id': 'bob-skyline',
#        'values': {'?principal': 'User::"bob"', '?resource': 'Photo::"skyline.jpg"'}}]},
#  {'id': 'policy1',                        # the second template, linked once
#   'id_annotation': 'photo-edit',
#   'slots': ['?principal', '?resource'],
#   'links': [
#       {'id': 'alice-edit-vacation',
#        'values': {'?principal': 'User::"alice"', '?resource': 'Photo::"vacation.jpg"'}}]}]

# the link ids live under each template's 'links'; pass one to without_linked
# to revoke that grant — returns a NEW handle; `linked` is unchanged
revoked = linked.without_linked("bob-skyline")
# revoked.templates()[0]['links'] now lists only alice-vacation

templates() is the one introspection call you need: each template's id / id_annotation / slots, and the links produced from it (each link's id and bound values). To act on a single grant — e.g. revoke it — read its id from links and pass it to without_linked(link_id). (Order within templates and links is not guaranteed.)

A linked policy carries two distinct labels, which matters when you read diagnostics. Its id is the new_id you gave it (alice-vacation) — unique to that link, and what appears in result.diagnostics.reasons when it fires. Its @id is inherited from the template (photo-access) — so every link of one template shares the same @id while keeping a distinct id. That is why a matched policy is identified by its id, not its @id: link a template for ten grants and ten policies share the @id, but each has its own id.

Because an unlinked template is inert — the authorizer only ever evaluates linked and static policies, never the templates themselves — you can parse a whole catalogue of templates into the base once and link only the ones a given request needs; the unused templates cost nothing at evaluation time.

Should you reach for templates? It depends on your policies and scale, and it's worth measuring rather than assuming. You would be right to suspect that at small scale and low complexity, generating a policy string per principal and parsing it is the more efficient choice — lower CPU and lower memory than linking, since parsing a short policy is cheaper than the per-link work of binding entity uids and cloning the rule's form. The two reach parity once the policy carries a few conditions, and templates pull ahead from there as the body gets richer — the rule is parsed and held once instead of copied per grant. In a quick build benchmark at 200 grants, a one-condition policy (like the subscription rule above) still favored generate-and-parse, a ~4-condition policy was about even, and a rich ~10-condition grant built about 2× faster and used about 2.5× less memory as a template. So profile your own policy shapes and grant counts, and take whichever approach gives the best result for your use case.

Partially authorizing a request with unknowns

Sometimes you can't fully evaluate a request up front. A resource entity may not be loaded from the database yet, the caller may not have picked a resource, or context may be filled in by a downstream service. is_authorized_partial evaluates whatever is known and returns residual policies for the parts that aren't.

is_authorized_partial returns either:

• Decision.Allow or Decision.Deny when the unknowns can't change the outcome, or
• Decision.NoDecision plus residual policies for the caller to re-evaluate once the unknowns are bound.

from cedarpy import is_authorized_partial, PartialAuthzResult, Decision

# Allow access to all principals when the resource is public
policies: str = 'permit(principal, action, resource) when { resource.public == true };'

# Resource entity hasn't been loaded from the database yet
entities: list = []

request = {
    "principal": 'User::"alice"',
    "action": 'Action::"view"',
    "resource": 'Photo::"photo1"',
    "context": {},
}

result: PartialAuthzResult = is_authorized_partial(request, policies, entities)

# Cedar can't decide yet: the policy depends on resource.public,
# but Photo::"photo1" hasn't been loaded.
assert result.decision == Decision.NoDecision

# Which entities Cedar needs you to load before it can decide:
assert result.diagnostics.unknown_entities == ['Photo::"photo1"']

# Which policies are still unresolved, identified by the auto-generated policy id:
assert result.diagnostics.nontrivial_residuals == ['policy0']

Now you can load the unknown entities and re-evaluate access with is_authorized to get a final decision:

from cedarpy import is_authorized

# Load Photo::"photo1" from the database and transform to its Cedar entity representation
entities = [
    {"uid": {"__entity": {"type": "Photo", "id": "photo1"}},
     "attrs": {"public": True}, "parents": []}
]

final_result = is_authorized(request, policies, entities) # new entities; same request and policies
assert final_result.decision == Decision.Allow

Note: A partial-eval result is not a final authorization decision. Always re-run is_authorized once unknowns are bound.

See the Partial Authorization Guide for edge cases, residual structure, and advanced patterns (SQL translation, Cedar text re-evaluation).

Validating policies against a schema

You can use validate_policies to validate Cedar policies against a schema before deploying them. Validation catches common mistakes like typos in entity types, invalid actions, type mismatches, and unsafe access to optional attributes—errors that would otherwise cause policies to silently fail at runtime.

This is particularly useful in CI/CD pipelines to catch policy errors before they reach production. See the Cedar validation documentation for details on what the validator checks.

Here's an example of basic use:

from cedarpy import validate_policies, Schema, ValidationResult

policies: str = "// a string containing Cedar policies"
schema = Schema.from_str(cedar_schema_text)  # or a JSON string, Cedar string, or dict

result: ValidationResult = validate_policies(policies, schema)

# so you can check validation passed like:
assert result.validation_passed

# or use ValidationResult in a boolean context
assert result  # True if validation passed

# and if validation fails, iterate over errors:
for error in result.errors:
    print(f"error: {error}")

The ValidationResult class provides the validation outcome and a list of ValidationError objects when validation fails.

See the unit tests for more examples of use and expected behavior.

Formatting Cedar policies

You can use format_policies to pretty-print Cedar policies according to convention.

from cedarpy import format_policies

policies: str = """
    permit(
        principal,
        action == Action::"edit",
        resource
    )
    when {
        resource.owner == principal
    };
"""

print(format_policies(policies))
# permit (
#   principal,
#   action == Action::"edit",
#   resource
# )
# when { resource.owner == principal };

Developing

You'll need a few things to get started:

• Python +3.9
• Rust and cargo

This project is built on the PyO3 and maturin projects. These projects are designed to enable Python to use Rust code and vice versa.

The most common development commands are in the Makefile

Create virtual env

First create a Python virtual environment for this project with: make venv-dev

In addition to creating a dedicated virtual environment, this will install cedar-py's dependencies.

If this works you should be able to run the following command:

maturin --help

Build and run cedar-py tests

Ensure the cedar-py virtual environment is active by sourcing it in your shell:

source venv-dev/bin/activate

Now run:

make quick

The make quick command will build the Rust source code with maturin and run the project's tests with pytest.

If all goes well, you should see output like:

(venv-dev) swedish-chef:cedar-py skuenzli$ make quick
Performing quick build
set -e ;\
	maturin develop ;\
	pytest
📦 Including license file "/path/to/cedar-py/LICENSE"
🔗 Found pyo3 bindings
🐍 Found CPython 3.9 at /path/to/cedar-py/venv-dev/bin/python
📡 Using build options features from pyproject.toml
Ignoring maturin: markers 'extra == "dev"' don't match your environment
Ignoring pip-tools: markers 'extra == "dev"' don't match your environment
Ignoring pytest: markers 'extra == "dev"' don't match your environment
💻 Using `MACOSX_DEPLOYMENT_TARGET=11.0` for aarch64-apple-darwin by default
   Compiling cedarpy v0.1.0 (/path/to/cedar-py)
    Finished dev [unoptimized + debuginfo] target(s) in 3.06s
📦 Built wheel for CPython 3.9 to /var/folders/k2/tnw8n1c54tv8nt4557pfx3440000gp/T/.tmpO6aj6c/cedarpy-0.1.0-cp39-cp39-macosx_11_0_arm64.whl
🛠 Installed cedarpy-0.1.0
================================================================================================ test session starts ================================================================================================
platform darwin -- Python 3.9.12, pytest-7.4.0, pluggy-1.2.0
rootdir: /path/to/cedar-py
configfile: pyproject.toml
testpaths: tests/unit
collected 10 items

tests/unit/test_authorize.py::AuthorizeTestCase::test_authorize_basic_ALLOW PASSED                                                                                                                            [ 10%]
tests/unit/test_authorize.py::AuthorizeTestCase::test_authorize_basic_DENY PASSED                                                                                                                             [ 20%]

... snip ... # a bunch of tests passing - please write more!
tests/unit/test_import_module.py::InvokeModuleTestFunctionTestCase::test_invoke_parse_test_policy PASSED                                                                                                      [100%]

================================================================================================ 10 passed in 0.51s =================================================================================================

Integration tests

This project supports validating correctness with official Cedar integration tests. To run those tests you'll need to retrieve the cedar-integration-tests data with:

make submodules

Then you can run:

make integration-tests

cedar-py currently passes all 74 tests defined in the example_use_cases, multi, ip, and decimal suites. The integration tests also validate policies against schemas when shouldValidate is set in the test definition. See test_cedar_integration_tests.py for details.

Corpus tests

The upstream cedar-integration-tests repository also ships a fuzzer-generated corpus (corpus-tests.tar.gz) — 7,462 test files containing 59,696 individual request cases. cedar-py passes all 59,696. Runs are kept in a separate target since the suite takes a few minutes:

make corpus-tests

The runner mirrors the leniency rules used by the upstream Rust runner (cedar-testing/tests/cedar-policy/corpus_tests.rs) and cedar-java's SharedIntegrationTests: reasons compared as a set, errors compared by count, and policy/schema parse failures (cedarpy's NoDecision) soft-pass when the fixture's expected decision is Deny. See test_cedar_corpus_tests.py for details.

Using locally-built artifacts

If you used make quick above, then a development build of the cedarpy module will already be installed in the virtual environment.

If you want to use your local cedarpy changes in another Python environment, you'll need to build a release with:

make release

The release process will build a wheel and output it into target/wheels/

Now you can install that file with pip, e.g.:

pip install --force-reinstall /path/to/cedar-py/target/wheels/ccedarpy-*.whl

Contributing

This project is in its early stages and contributions are welcome. Please check the project's GitHub issues for work we've already identified.

Some ways to contribute are:

• Use the project and report experience and issues
• Document usage and limitations
• Enhance the library with additional functionality you need
• Add test cases, particularly those from cedar-integration-tests

You can reach people interested in this project in the #cedar-py channel of the Cedar Policy Slack workspace.