magic-wormhole/docs/api.md

# Magic-Wormhole

This library provides a mechanism to securely transfer small amounts
of data between two computers. Both machines must be connected to the
internet, but they do not need to have public IP addresses or know how to
contact each other ahead of time.

Security and connectivity is provided by means of an "wormhole code": a short
string that is transcribed from one machine to the other by the users at the
keyboard. This works in conjunction with a baked-in "rendezvous server" that
relays information from one machine to the other.

The "Wormhole" object provides a secure record pipe between any two programs
that use the same wormhole code (and are configured with the same application
ID and rendezvous server). Each side can send multiple messages to the other,
but the encrypted data for all messages must pass through (and be temporarily
stored on) the rendezvous server, which is a shared resource. For this
reason, larger data (including bulk file transfers) should use the Transit
class instead. The Wormhole object has a method to create a Transit object
for this purpose. In the future, Transit will be deprecated, and this
functionality will be incorporated directly as a "dilated wormhole".

A quick example:

```python
import wormhole
from twisted.internet.defer import inlineCallbacks

@inlineCallbacks
def go():
    w = wormhole.create(appid, relay_url, reactor)
    w.generate_code()
    code = yield w.when_code()
    print "code:", code
    w.send(b"outbound data")
    inbound = yield w.when_received()
    yield w.close()
```

## Modes

The API comes in two flavors: Delegated and Deferred. Controlling the
Wormhole and sending data is identical in both, but they differ in how
inbound data and events are delivered to the application.

In Delegated mode, the Wormhole is given a "delegate" object, on which
certain methods will be called when information is available (e.g. when the
code is established, or when data messages are received). In Deferred mode,
the Wormhole object has methods which return Deferreds that will fire at
these same times.

Delegated mode:

```python
class MyDelegate:
    def wormhole_got_code(self, code):
        print("code: %s" % code)
    def wormhole_received(self, data): # called for each message
        print("got data, %d bytes" % len(data))

w = wormhole.create(appid, relay_url, reactor, delegate=MyDelegate())
w.generate_code()
```

Deferred mode:

```python
w = wormhole.create(appid, relay_url, reactor)
w.generate_code()
def print_code(code):
    print("code: %s" % code)
w.when_code().addCallback(print_code)
def received(data):
    print("got data, %d bytes" % len(data))
w.when_received().addCallback(received) # gets exactly one message
```

## Application Identifier

Applications using this library must provide an "application identifier", a
simple string that distinguishes one application from another. To ensure
uniqueness, use a domain name. To use multiple apps for a single domain,
append a URL-like slash and path, like `example.com/app1`. This string must
be the same on both clients, otherwise they will not see each other. The
invitation codes are scoped to the app-id. Note that the app-id must be
unicode, not bytes, so on python2 use `u"appid"`.

Distinct app-ids reduce the size of the connection-id numbers. If fewer than
ten Wormholes are active for a given app-id, the connection-id will only need
to contain a single digit, even if some other app-id is currently using
thousands of concurrent sessions.

## Rendezvous Servers

The library depends upon a "rendezvous server", which is a service (on a
public IP address) that delivers small encrypted messages from one client to
the other. This must be the same for both clients, and is generally baked-in
to the application source code or default config.

This library includes the URL of a public rendezvous server run by the
author. Application developers can use this one, or they can run their own
(see the `wormhole-server` command and the `src/wormhole/server/` directory)
and configure their clients to use it instead. This URL is passed as a
unicode string. Note that because the server actually speaks WebSockets, the
URL starts with `ws:` instead of `http:`.

## Wormhole Parameters

All wormholes must be created with at least three parameters:

* `appid`: a (unicode) string
* `relay_url`: a (unicode) string
* `reactor`: the Twisted reactor object

In addition to these three, the `wormhole.create()` function takes several
optional arguments:

* `delegate`: provide a Delegate object to enable "delegated mode", or pass
  None (the default) to get "deferred mode"
* `journal`: provide a Journal object to enable journaled mode. See
  journal.md for details. Note that journals only work with delegated mode,
  not with deferred mode.
* `tor_manager`: to enable Tor support, create a `wormhole.TorManager`
  instance and pass it here. This will hide the client's IP address by
  proxying all connections (rendezvous and transit) through Tor. It also
  enables connecting to Onion-service transit hints, and (in the future) will
  enable the creation of Onion-services for transit purposes.
* `timing`: this accepts a DebugTiming instance, mostly for internal
  diagnostic purposes, to record the transmit/receive timestamps for all
  messages. The `wormhole --dump-timing=` feature uses this to build a
  JSON-format data bundle, and the `misc/dump-timing.py` tool can build a
  scrollable timing diagram from these bundles.
* `welcome_handler`: this is a function that will be called when the
  Rendezvous Server's "welcome" message is received. It is used to display
  important server messages in an application-specific way.
* `app_versions`: this can accept a dictionary (JSON-encodable) of data that
  will be made available to the peer via the `got_version` event. This data
  is delivered before any data messages, and can be used to indicate peer
  capabilities.

## Code Management

Each wormhole connection is defined by a shared secret "wormhole code". These
codes can be generated offline (by picking a unique number and some secret
words), but are more commonly generated by whoever creates the first
wormhole. In the "bin/wormhole" file-transfer tool, the default behavior is
for the sender to create the code, and for the receiver to type it in.

The code is a (unicode) string in the form `NNN-code-words`. The numeric
"NNN" prefix is the "channel id" or "nameplate", and is a short integer
allocated by talking to the rendezvous server. The rest is a
randomly-generated selection from the PGP wordlist, providing a default of 16
bits of entropy. The initiating program should display this code to the user,
who should transcribe it to the receiving user, who gives it to their local
Wormhole object by calling `set_code()`. The receiving program can also use
`input_code()` to use a readline-based input function: this offers tab
completion of allocated channel-ids and known codewords.

The Wormhole object has three APIs for generating or accepting a code:

* `w.generate_code(length=2)`: this contacts the Rendezvous Server, allocates
  a short numeric nameplate, chooses a configurable number of random words,
  then assembles them into the code
* `w.set_code(code)`: this accepts the code as an argument
* `helper = w.input_code()`: this facilitates interactive entry of the code,
  with tab-completion. The helper object has methods to return a list of
  viable completions for whatever portion of the code has been entered so
  far. A convenience wrapper is provided to attach this to the `rlcompleter`
  function of libreadline.

No matter which mode is used, the `w.when_code()` Deferred (or
`delegate.wormhole_got_code(code)` callback) will fire when the code is
known. `when_code` is clearly necessary for `generate_code`, since there's no
other way to learn what code was created, but it may be useful in other modes
for consistency.

The code-entry Helper object has the following API:

* `update_nameplates()`: requests an updated list of nameplates from the
  Rendezvous Server. These form the first portion of the wormhole code (e.g.
  "4" in "4-purple-sausages"). Note that they are unicode strings (so "4",
  not 4). The Helper will get the response in the background, and calls to
  `complete_nameplate()` after the response will use the new list.
* `completions = h.complete_nameplate(prefix)`: returns (synchronously) a
  list of suffixes for the given nameplate prefix. For example, if the server
  reports nameplates 1, 12, 13, 24, and 170 are in use,
  `complete_nameplate("1")` will return `["", "2", "3", "70"]`. Raises
  `AlreadyClaimedNameplateError` if called after `h.claim_nameplate`.
* `d = h.claim_nameplate(nameplate)`: accepts a string with the chosen
  nameplate. May only be called once, after which `OnlyOneNameplateError` is
  raised. Returns a Deferred that fires (with None) when the nameplate's
  wordlist is known (which happens after the nameplate is claimed, requiring
  a roundtrip to the server).
* `completions = h.complete_words(prefix)`: return (synchronously) a list of
  suffixes for the given words prefix. The possible completions depend upon
  the wordlist in use for the previously-claimed nameplate, so calling this
  before `claim_nameplate` will raise `MustClaimNameplateFirstError`. Given a
  prefix like "su", this returns a list of strings which are appropriate to
  append to the prefix (e.g. `["pportive", "rrender", "spicious"]`, for
  expansion into "supportive", "surrender", and "suspicious". The prefix
  should not include the nameplate, but *should* include whatever words and
  hyphens have been typed so far (the default wordlist uses alternate lists,
  where even numbered words have three syllables, and odd numbered words have
  two, so the completions depend upon how many words are present, not just
  the partial last word). E.g. `complete_words("pr")` will return
  `["ocessor", "ovincial", "oximate"]`, while `complete_words("opulent-pr")`
  will return `["eclude", "efer", "eshrunk", "inter", "owler"]`.
  If the wordlist is not yet known (i.e. the Deferred from `claim_nameplate`
  has not yet fired), this returns an empty list. It will also return an
  empty list if the prefix is complete (the last word matches something in
  the completion list, and there are no longer extension words), although the
  code may not yet be complete if there are additional words. The completions
  will never include a hyphen: the UI frontend must supply these if desired.
* `h.submit_words(words)`: call this when the user is finished typing in the
  code. It does not return anything, but will cause the Wormhole's
  `w.when_code()` (or corresponding delegate) to fire, and triggers the
  wormhole connection process. This accepts a string like "purple-sausages",
  without the nameplate. It must be called after `h.claim_nameplate()` or
  `MustClaimNameplateFirstError` will be raised.

The `rlcompleter` wrapper is a function that knows how to use the code-entry
helper to do tab completion of wormhole codes:

```python
from wormhole import create, rlcompleter_helper
w = create(appid, relay_url, reactor)
rlcompleter_helper("Wormhole code:", w.input_code())
d = w.when_code()
```

This helper runs python's `rawinput()` function inside a thread, since
`rawinput()` normally blocks.

The two machines participating in the wormhole setup are not distinguished:
it doesn't matter which one goes first, and both use the same Wormhole
constructor function. However if `w.generate_code()` is used, only one side
should use it.

## Offline Codes

In most situations, the "sending" or "initiating" side will call
`w.generate_code()` and display the resulting code. The sending human reads
it and speaks, types, performs charades, or otherwise transmits the code to
the receiving human. The receiving human then types it into the receiving
computer, where it either calls `w.set_code()` (if the code is passed in via
argv) or `w.input_code()` (for interactive entry).

Usually one machine generates the code, and a pair of humans transcribes it
to the second machine (so `w.generate_code()` on one side, and `w.set_code()`
or `w.input_code()` on the other). But it is also possible for the humans to
generate the code offline, perhaps at a face-to-face meeting, and then take
the code back to their computers. In this case, `w.set_code()` will be used
on both sides. It is unlikely that the humans will restrict themselves to a
pre-established wordlist when manually generating codes, so the completion
feature of `w.input_code()` is not helpful.

When the humans create an invitation code out-of-band, they are responsible
for choosing an unused channel-ID (simply picking a random 3-or-more digit
number is probably enough), and some random words. Dice, coin flips, shuffled
cards, or repeated sampling of a high-resolution stopwatch are all useful
techniques. The invitation code uses the same format either way: channel-ID,
a hyphen, and an arbitrary string. There is no need to encode the sampled
random values (e.g. by using the Diceware wordlist) unless that makes it
easier to transcribe: e.g. rolling 6 dice could result in a code like
"913-166532", and flipping 16 coins could result in "123-HTTHHHTTHTTHHTHH".

## Verifier

For extra protection against guessing attacks, Wormhole can provide a
"Verifier". This is a moderate-length series of bytes (a SHA256 hash) that is
derived from the supposedly-shared session key. If desired, both sides can
display this value, and the humans can manually compare them before allowing
the rest of the protocol to proceed. If they do not match, then the two
programs are not talking to each other (they may both be talking to a
man-in-the-middle attacker), and the protocol should be abandoned.

Once retrieved, you can turn this into hex or Base64 to print it, or render
it as ASCII-art, etc. Once the users are convinced that `verify()` from both
sides are the same, call `send()` to continue the protocol. If you call
`send()` before `verify()`, it will perform the complete protocol without
pausing.

## Events

As the wormhole connection is established, several events may be dispatched
to the application. In Delegated mode, these are dispatched by calling
functions on the delegate object. In Deferred mode, the application retrieves
Deferred objects from the wormhole, and event dispatch is performed by firing
those Deferreds.

* got_code (`yield w.when_code()` / `dg.wormhole_code(code)`): fired when the
  wormhole code is established, either after `w.generate_code()` finishes the
  generation process, or when the Input Helper returned by `w.input_code()`
  has been told `h.set_words()`, or immediately after `w.set_code(code)` is
  called. This is most useful after calling `w.generate_code()`, to show the
  generated code to the user so they can transcribe it to their peer.
* verified (`verifier = yield w.when_verified()` /
  `dg.wormhole_verified(verifier)`: fired when the key-exchange process has
  completed and a valid VERSION message has arrived. The "verifier" is a byte
  string with a hash of the shared session key; clients can compare them
  (probably as hex) to ensure that they're really talking to each other, and
  not to a man-in-the-middle. When `got_verifier` happens, this side knows
  that *someone* has used the correct wormhole code; if someone used the
  wrong code, the VERSION message cannot be decrypted, and the wormhole will
  be closed instead.
* version (`yield w.when_version()` / `dg.wormhole_version(version)`:
  fired when the VERSION message arrives from the peer. This fires at the
  same time as `verified`, but delivers the "app_versions" data (passed into
  `wormhole.create`) instead of the verifier string.
* received (`yield w.when_received()` / `dg.wormhole_received(data)`: fired
  each time a data message arrives from the peer, with the bytestring that
  the peer passed into `w.send(data)`.
* closed (`yield w.close()` / `dg.wormhole_closed(result)`: fired when
  `w.close()` has finished shutting down the wormhole, which means all
  nameplates and mailboxes have been deallocated, and the WebSocket
  connection has been closed. This also fires if an internal error occurs
  (specifically WrongPasswordError, which indicates that an invalid encrypted
  message was received), which also shuts everything down. The `result` value
  is an exception (or Failure) object if the wormhole closed badly, or a
  string like "happy" if it had no problems before shutdown.

## Sending Data

The main purpose of a Wormhole is to send data. At any point after
construction, callers can invoke `w.send(data)`. This will queue the message
if necessary, but (if all goes well) will eventually result in the peer
getting a `received` event and the data being delivered to the application.

Since Wormhole provides an ordered record pipe, each call to `w.send` will
result in exactly one `received` event on the far side. Records are not
split, merged, dropped, or reordered.

Each side can do an arbitrary number of `send()` calls. The Wormhole is not
meant as a long-term communication channel, but some protocols work better if
they can exchange an initial pair of messages (perhaps offering some set of
negotiable capabilities), and then follow up with a second pair (to reveal
the results of the negotiation). The Rendezvous Server does not currently
enforce any particular limits on number of messages, size of messages, or
rate of transmission, but in general clients are expected to send fewer than
a dozen messages, of no more than perhaps 20kB in size (remember that all
these messages are temporarily stored in a SQLite database on the server). A
future version of the protocol may make these limits more explicit, and will
allow clients to ask for greater capacity when they connect (probably by
passing additional "mailbox attribute" parameters with the
`allocate`/`claim`/`open` messages).

For bulk data transfer, see "transit.md", or the "Dilation" section below.

## Closing

When the application is done with the wormhole, it should call `w.close()`,
and wait for a `closed` event. This ensures that all server-side resources
are released (allowing the nameplate to be re-used by some other client), and
all network sockets are shut down.

In Deferred mode, this just means waiting for the Deferred returned by
`w.close()` to fire. In Delegated mode, this means calling `w.close()` (which
doesn't return anything) and waiting for the delegate's `wormhole_closed()`
method to be called.

## Serialization

(this section is speculative: this code has not yet been written)

Wormhole objects can be serialized. This can be useful for apps which save
their own state before shutdown, and restore it when they next start up
again.


The `w.serialize()` method returns a dictionary which can be JSON encoded
into a unicode string (most applications will probably want to UTF-8 -encode
this into a bytestring before saving on disk somewhere).

To restore a Wormhole, call `wormhole.from_serialized(data, reactor,
delegate)`. This will return a wormhole in roughly the same state as was
serialized (of course all the network connections will be disconnected).

Serialization only works for delegated-mode wormholes (since Deferreds point
at functions, which cannot be serialized easily). It also only works for
"non-dilated" wormholes (see below).

To ensure correct behavior, serialization should probably only be done in
"journaled mode". See journal.md for details.

If you use serialization, be careful to never use the same partial wormhole
object twice.

## Bytes, Strings, Unicode, and Python 3

All cryptographically-sensitive parameters are passed as bytes ("str" in
python2, "bytes" in python3):

* verifier string
* data in/out
* transit records in/out

Other (human-facing) values are always unicode ("unicode" in python2, "str"
in python3):

* wormhole code
* relay URL
* transit URLs
* transit connection hints (e.g. "host:port")
* application identifier
* derived-key "purpose" string: `w.derive_key(PURPOSE, LENGTH)`

## Full API list

action                     | Deferred-Mode        | Delegated-Mode
-------------------------- | -------------------- | --------------
w.generate_code(length=2)  |                      |
w.set_code(code)           |                      |
h=w.input_code()           |                      |
                           |  d=w.when_code()     | dg.wormhole_code(code)
                           |  d=w.when_verified() | dg.wormhole_verified(verifier)
                           |  d=w.when_version()  | dg.wormhole_version(version)
w.send(data)               |                      |
                           |  d=w.when_received() | dg.wormhole_received(data)
key=w.derive_key(purpose, length)  |              |
w.close()                  |                      | dg.wormhole_closed(result)
                           |  d=w.close()         |


## Dilation

(this section is speculative: this code has not yet been written)

In the longer term, the Wormhole object will incorporate the "Transit"
functionality (see transit.md) directly, removing the need to instantiate a
second object. A Wormhole can be "dilated" into a form that is suitable for
bulk data transfer.

All wormholes start out "undilated". In this state, all messages are queued
on the Rendezvous Server for the lifetime of the wormhole, and server-imposed
number/size/rate limits apply. Calling `w.dilate()` initiates the dilation
process, and success is signalled via either `d=w.when_dilated()` firing, or
`dg.wormhole_dilated()` being called. Once dilated, the Wormhole can be used
as an IConsumer/IProducer, and messages will be sent on a direct connection
(if possible) or through the transit relay (if not).

What's good about a non-dilated wormhole?:

* setup is faster: no delay while it tries to make a direct connection
* survives temporary network outages, since messages are queued
* works with "journaled mode", allowing progress to be made even when both
  sides are never online at the same time, by serializing the wormhole

What's good about dilated wormholes?:

* they support bulk data transfer
* you get flow control (backpressure), and provide IProducer/IConsumer
* throughput is faster: no store-and-forward step

Use non-dilated wormholes when your application only needs to exchange a
couple of messages, for example to set up public keys or provision access
tokens. Use a dilated wormhole to move large files.

Dilated wormholes can provide multiple "channels": these are multiplexed
through the single (encrypted) TCP connection. Each channel is a separate
stream (offering IProducer/IConsumer)

To create a channel, call `c = w.create_channel()` on a dilated wormhole. The
"channel ID" can be obtained with `c.get_id()`. This ID will be a short
(unicode) string, which can be sent to the other side via a normal
`w.send()`, or any other means. On the other side, use `c =
w.open_channel(channel_id)` to get a matching channel object.

Then use `c.send(data)` and `d=c.when_received()` to exchange data, or wire
them up with `c.registerProducer()`. Note that channels do not close until
the wormhole connection is closed, so they do not have separate `close()`
methods or events. Therefore if you plan to send files through them, you'll
need to inform the recipient ahead of time about how many bytes to expect.