#auth #token #delegation #authorization #decentralized #verification


an authentication/authorization token with decentralized verification and offline delegation

8 releases (5 breaking)

0.6.0 Sep 11, 2020
0.5.2 Jul 29, 2020
0.5.1 Jun 17, 2020
0.5.0 Feb 13, 2020
0.1.0 Nov 26, 2019

#7 in Authentication

48 downloads per month



Biscuit authentication and authorization token

Biscuit is an authorization token for microservices architectures with the following properties:

  • decentralized validation: any node could validate the token only with public information;
  • offline delegation: a new, valid token can be created from another one by attenuating its rights, by its holder, without communicating with anyone;
  • capabilities based: authorization in microservices should be tied to rights related to the request, instead of relying to an identity that might not make sense to the verifier;
  • flexible rights managements: the token uses a logic language to specify attenuation and add bounds on ambient data;
  • small enough to fit anywhere (cookies, etc).

Non goals:

  • This is not a new authentication protocol. Biscuit tokens can be used as opaque tokens delivered by other systems such as OAuth.
  • Revocation: while tokens come with expiration dates, revocation requires external state management.


In this example we will see how we can create a token, add some caveats, serialize and deserialize a token, append more caveats, and validate those caveats in the context of a request:

extern crate rand;
extern crate biscuit_auth as biscuit;

use biscuit::{crypto::KeyPair, token::{Biscuit, verifier::Verifier, builder::*}, error};

fn main() -> Result<(), error::Token> {
  let mut rng = rand::thread_rng();

  // let's generate the root key pair. The root public key will be necessary
  // to verify the token
  let root = KeyPair::new(&mut rng);
  let public_key = root.public();

  // creating a first token
  let token1 = {
    // the first block of the token is the authority block. It contains global
    // information like which operation types are available
    let mut builder = Biscuit::builder(&mut rng, &root);

    // let's define some access rights
    // every fact added to the authority block must have the authority fact
    builder.add_authority_fact("right(#authority, \"/a/file1.txt\", #read)")?;
    builder.add_authority_fact("right(#authority, \"/a/file1.txt\", #write)")?;
    builder.add_authority_fact("right(#authority, \"/a/file2.txt\", #read)")?;
    builder.add_authority_fact("right(#authority, \"/b/file3.txt\", #write)")?;

    // we can now create the token
    let biscuit = builder.build()?;
    println!("biscuit (authority): {}", biscuit.print());


  // this token is only 266 bytes, holding the authority data and the signature
  assert_eq!(token1.len(), 266);

  // now let's add some restrictions to this token
  // we want to limit access to `/a/file1.txt` and to read operations
  let token2 = {
    // the token is deserialized, the signature is verified
    let deser = Biscuit::from(&token1)?;

    let mut builder = deser.create_block();

    // caveats are implemented as logic rules. If the rule produces something,
    // the caveat is successful
      // the rule's name
      // the "head" of the rule, defining the kind of result that is produced
      // here we require the presence of a "resource" fact with the "ambient" tag
      // (meaning it is provided by the verifier)
        pred("resource", &[s("ambient"), string("/a/file1.txt")]),
        // we restrict to read operations
        pred("operation", &[s("ambient"), s("read")]),

    // the previous caveat could also be written like this
    // builder.add_caveat("caveat(#resource) <- resource(#ambient, \"/a/file1.txt\"), operation(#ambient, #read)")?;

    let keypair = KeyPair::new(&mut rng);
    // we can now create a new token
    let biscuit = deser.append(&mut rng, &keypair, builder.build())?;
    println!("biscuit (authority): {}", biscuit.print());


  // this new token fits in 402 bytes
  assert_eq!(token2.len(), 402);

  /************** VERIFICATION ****************/

  // let's deserialize the token:
  let biscuit2 = Biscuit::from(&token2)?;

  // let's define 3 verifiers (corresponding to 3 different requests):
  // - one for /a/file1.txt and a read operation
  // - one for /a/file1.txt and a write operation
  // - one for /a/file2.txt and a read operation

  let mut v1 = biscuit2.verify(public_key)?;
  // we will check that the token has the corresponding right
  v1.add_rule("read_right(#read_right) <- right(#authority, \"/a/file1.txt\", #read)");

  let mut v2 = biscuit2.verify(public_key)?;
  v2.add_rule("write_right(#write_right) <- right(#authority, \"/a/file1.txt\", #write)");

  let mut v3 = biscuit2.verify(public_key)?;
  v3.add_rule("read_right(#read_right) <- right(#authority, \"/a/file2.txt\", #read)");

  // the token restricts to read operations:
  // the second verifier requested a read operation
  // the third verifier requests /a/file2.txt




A Biscuit token is made with a list of blocks defining data and caveats that must be validated upon reception with a request. Any failed caveat will invalidate the entire token.

If you hold a valid token, it is possible to add a new block to restrict further the token, like limiting access to one particular resource, or adding a short expiration date. This will generate a new, valid token. This can be done offline, without asking the original token creator.

On the other hand, if a block is modified or removed, the token will fail the cryptographic signature verification.


Biscuit tokens get inspiration from macaroons and JSON Web Tokens, reproducing useful features from both:

  • offline delegation like macaroons
  • based on public key cryptography like JWT, so any application holding the root public key can verify a token (while macaroons are based on a root shared secret)

A logic language for caveats: Datalog with constraints

We rely on a modified version of Datalog, that can represent complex behaviours in a compact form, and add flexible constraints on data.

Here are examples of caveats that can be implemented with that language:

  • valid if the requested resource is "file.txt" and the operation is "read"
  • valid if current time is before January 1st 2030, 00h00mn00s UTC
  • source IP is in set [,]
  • resource matches prefix "/home/biscuit/data/"
  • But it can also combine into more complex patterns, like: right is read if user has read or user is member of organisation and organisation has read right or other user with read right has delegated to user.

Like Datalog, this language is based around facts and rules, but with some slight modifications:

  • an authority fact starts with the #authority symbol. It can only be added in the authority block (or generated from rules in the authority rules). It provides the basic authorization data, like which rights exist
  • an ambient fact starts with the #ambient symbol. It can only be provided by the verifier. It gives information on the current request, like which resource is accessed or the current time
  • Blocks can provide facts but they cannot be authority or ambient facts. They contain rules that use facts from the current block, or from the authority and ambient contexts. If all rules in a block succeed, the block is validated.

A caveat rule requires the presence of one or more facts, and can have additional constraints on these facts (the constraints are implemented separately to simplify the language implementation: among other things, it avoids implementing negation). It is possible to create rules like these ones:

  • caveat = resource("file1")
  • caveat = resource(0?) & owner("user1", 0?) // the 0? represents a "hole" that must be filled with the correct value
  • caveat = time(0?) | 0? < 2019-02-05T23:00:00Z // expiration date
  • application(0?) & operation(1?) &user(2?) & & right(app, 0?, 1?) & owner(2?, 0?) & credit(2?, 3?) | 3? > 0 // verifies that the user owns the applications, the application has the right on the operation, there's a credit information for the operation, and the credit is larger than 0

Symbols and symbol tables

To reduce the size of tokens, the language supports a data type called "symbol". A symbol is a string that we can refer to with a number, an index in the symbol table that is carried with the token. Symbols can be checked for equality, or presence in a set, but lack the other constraints on strings like prefix or suffix matching.

They can be used for pretty printing of a fact or rule. As an example, with a table containing ["resource", "operation", "read", "caveat1"], we could have the following rule: #4 <- #0("file.txt") & #1(#2)that would be printed ascaveat1 <- resoucr("file.txt") & operation(read)`

biscuit implementations come with a default symbol table to avoid transmitting frequent values with every token.


Licensed under Apache License, Version 2.0, (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)


Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be licensed as above, without any additional terms or conditions.


~136K SLoC