Writing Policies

Policies are the source of truth for the authorization logic used to evaluate queries in oso. As a reminder: oso policies are written in a declarative language called Polar. There is a full Polar Syntax guide which you can use as a reference of all the available syntax, but here we’ll give an overview of getting started with writing policies.

The syntax might feel a bit foreign at first, but fear not: almost anything you can express in imperative code can equally be expressed in Polar — often more concisely and closer to how you might explain the logic in natural language.


Policies are stored in Polar files (extension .polar), which are loaded into the authorization engine using the oso Libraries.

Rule Basics

Policies are made up of rules. Each rule defines a statement that is either true or false. oso answers queries by evaluating rules that match the query name and parameters. Let’s take a basic allow rule as an example:

allow(actor, action, resource) if ...

When we use is_allowed() (or equivalent), we are making a query that asks oso to evaluate all rules that match (a) on the rule name "allow", and (b) on all the inputs.

In the rule above, actor, action, and resource are simply the parameter names, i.e. they are variables that will match anything.

But if we replace action with a concrete type:

allow(actor, "read", resource);

the rule will now only be evaluated if the second input exactly matches the string "read".

if Statements

There are several ways to represent imperative if logic in Polar.

In a Rule Body

The most common way to write an if statement in Polar is to add a body to a rule. The following rule allows any actor to approve any expense report:

allow(_actor, "approve", _report);

To restrict the rule such that only administrators may approve any expense report, we can add a body:

allow(actor, "approve", _report) if
    actor.is_admin = true;

To express multiple truth conditions (e.g., if A or B, then...), we can either create multiple rules…

allow(actor, "approve", _report) if
    actor.is_admin = true;

allow(actor, "approve", _report) if
    actor.title = "CFO";

…or we can use Polar’s Disjunction (or) operator (OR) to combine the conditions in a single rule body:

allow(actor, "approve", _report) if
    actor.is_admin = true
    or actor.title = "CFO";


In these rules we declared some variables with leading underscores (_report, _actor). A leading underscore indicates that the variable will only be used once (Polar does not distinguish between definition and use). These variables are called singleton variables, and will match any value. To help prevent errors, a warning will be emitted if a singleton variable is not preceded by an underscore.

As Specializers in a Rule Head

Given the following application class structure…

class User:

class Admin(User):

…we can modify our original bodiless rule to only allow Admin users to approve any expense report by adding a specializer to the rule head:

allow(_actor: Admin, "approve", _report);

The rule will fail when evaluated on a regular User and succeed when evaluated on an Admin, encoding an implicit if Admin condition.

This is another example of the rule matching process: instead of matching against a concrete value, we are instead checking to make sure the type of the input matches the expected type - in this case, an Admin.


Try to use type specializers as often as possible. It will help make sure you don’t accidentally allow access to an unrelated resource which happens to have matching fields.

Combining Rules

Rules can be thought of as equivalent to methods in imperative programming. And the same idea should be applied when writing policies: any pieces of logic that you want to reuse throughout a policy can be extracted out into a new rule.

The benefit of this is (a) it makes it easier to keep logic consistent throughout, and (b) it often results in much more readable policy.

Take the following example. We want a rule saying that accountants can read expenses. Our initial version might look like:

allow(user: User, "read", expense: Expense) if
  user.role = "accountant";

This would be fine, but if, for example, we wanted to allow the CFO to do whatever an accountant can do, we would need to duplicate all the rules. Or if we want to change how an application determines roles we would need to change all locations using this.

So instead, we can refactor the role check into its own rule:

allow(user: User, "read", expense: Expense) if
  role(user, "accountant");

role(user, role_name) if user.role = role_name;

The role(user, "accountant") is yet another example of matching happening in Polar. Any time a rule body contains a predicate like this, it is performing another query. I.e. it will try and find all matching rules called “role” with two inputs.

You can also either use the The REPL or the oso.query_rule method to interact with this directly. For example:

from oso import Oso

class User:
    def __init__(self, name, role):
        self.name = name
        self.role = role

oso = Oso()
oso.load_str("role(user, role_name) if user.role = role_name;")

alice = User("alice", "accountant")
assert list(oso.query_rule("role", alice, "accountant"))


Try setting the POLAR_LOG environment variable before executing a polar query to see a trace of how the query is evaluated:

$ POLAR_LOG=1 python user.py
[debug]   QUERY: role(<__main__.User object at 0x105da6190>, "accountant"), BINDINGS: {}
[debug]       role(user, role_name) if user.role = role_name;
[debug]     RULE: role(user, role_name) if user.role = role_name;
[debug]       QUERY: .(_user_5, "role", _value_1_7) and _value_1_7 = _role_name_6, BINDINGS: {_role_name_6 = "accountant", _user_5 = <__main__.User object at 0x105da6190>}
[debug]         QUERY: .(_user_5, "role", _value_1_7), BINDINGS: {_user_5 = <__main__.User object at 0x105da6190>}
[debug]           LOOKUP: <__main__.User object at 0x105da6190>.role()
[debug]           => "accountant"
[debug]         QUERY: _value_1_7 = _role_name_6, BINDINGS: {_role_name_6 = "accountant", _value_1_7 = "accountant"}
[debug]   BACKTRACK
[debug]           LOOKUP: <__main__.User object at 0x105da6190>.role()
[debug]           => No more results.
[debug]           BACKTRACK
[debug]           HALT


We covered some of the basics of policies, how to represent conditional logic, and the core idea of matching.

What’s next

  • Interested in understanding more about leveraging application types for specializers and attribute lookups? Continue on to Application Types.

  • To see more policy examples—including guides oriented around specific models like role-based or attribute-based access control—see Policy Examples.

  • To continue learning policy syntax, go on to Polar Syntax.