DISCLAIMER: Although I also work with solid in my day job at inrupt, I wrote this guide in my spare time. There are places where I chose one side of a complex unresolved discussion from an unhosted web apps perspective, so subject to change!
How unhosted web apps can interact with solid pods
Solid is similar to remoteStorage in that it allows apps and services (including unhosted web apps) to store the user's data under the user's control. Where it differs from remoteStorage is mainly in its smart use of linked data principles, the existence of a PATCH verb, and its versatile Access Control system.
This article describes how an app can interact with a user's Solid pod.
A WebId is a URL that uniquely identifies a user. Using WebId-OIDC, the app can obtain an id token which proves that the user currently interacting with the app controls that WebId. From this id token, a bearer token can be derived. The exact way in which this happens will change in 2020. The old way is described in the WebId-OIDC spec, and the new way has not been fully described yet. Solid apps should support both old and new identity providers, as well as old and new storage providers. But luckily, as an app developer you have tools on your side. You can follow for instance this excellent tutorial and rely on the Solid Auth Client library to handle the WebId discovery and authentication.
Solid Auth Client will take care of opening a popup window where the user can find their identity provider from a list, or type in their WebId as a URL. When the token comes back to the app, Solid Auth Client will harvest it from there, and use it to create a bearer token. You can then use SolidAuth.fetch instead of the browser's built-in fetch function, and it will behave just like the normal fetch, but adding the appropriate Authorization header to each request. Popular data access libraries for Solid, like Tripledoc, RDFLib, and LDflex, will automatically use SolidAuth.fetch as well.
Web of Personal Data
Although you could store any data on the user's pod, by convention, personal data is organized in a particular way on a Solid pod. The starting point for the web of data around the currently interacting user is their profile document, which is always hosted at the user's WebId, and is always publicly readable. A WebId usually contains a hash, for instance https://michielbdejong.inrupt.net/profile/card#me. The document will be an RDF data source, giving you (subject, predicate, object) triples in for instance text/turtle or json-ld representation. Turtle is pretty easy to read (to see an example, click on the </> icon on my Solid profile); at the top is usually a list of prefixes which end in :, so for instance ldp: is an abbreviation for 'https://www.w3.org/ns/ldp#'. So when you see 'ldp:inbox', it means 'https://www.w3.org/ns/ldp#inbox'. You pretty much always want to start by fetching the profile document of the currently logged in user. From there, you can discover the ldp:inbox link to the user's inbox, solid:privateTypeIndex, solid:publicTypeIndex, and other data.
The inbox is a folder to which anybody can POST, but only the user themselves can read. It is useful for sending the user a message. Some domain-specific inboxes also exist in Solid, for instance for playing Tic-Tac-Toe and for friend requests, messages can be sent either to the user's global inbox, or to the domain-specific one if it exists.
The public and private type indexes lists zero or more solid:TypeRegistrations, with solid:forClass triples pointing to public and private documents that describe an instance of a specific RDF type. For instance, the public type index might link to a user's public list of bookmarks.
This system makes Solid more flexible than for instance remoteStorage, where private bookmarks always need to be at /bookmarks, and public ones always need to be at /public/bookmarks. On the other hand, it makes it harder to do scoped access control. As we'll see in the 'ACL documents' chapter further on, it's possible to give an app read, append, write, and/or control access to specific subtrees of the storage. This is generally done by a launcher app, which acts as an auth server, in that it has full access itself, but edits the access control lists when you're about to use an app that is only trusted for a certain data domain. Using its root permissions, the launcher app can also create folders and documents if they're missing, and add links to them in for instance the private or public type index of the user - things a domain-specific app is probably not allowed to do itself.
A full overview of data that you might find on a user's pod, and how to discover it, is in the client-client spec which is currently under construction, but a good draft so far can be found in the data conventions of the Solid databrowser.
Reading and writing data is pretty similar to what you would expect if you know a bit about HTTP and you know how it works for remoteStorage. To read, the usual HEAD and GET verbs can be used. Data is stored hierarchically with files in folders, which are formally defined as LDP Resources and LDP BasicContainers. There are 4 types of resources: containers (their URL ends in a forward slash), ACL documents, other RDF sources, and non-RDF resources. RDF sources (i.e. Containers, ACL documents, and other RDF sources) can be thought of as not really tied to a text/turtle or json-ld representation, since the server is required to convert between at least those two, through content negotiation. So when you retrieve it requesting a Turtle representation, it's unknown to you whether it was also uploaded in Turtle, or in one of RDF's other representations. In the line of this way of thinking about RDF sources separately from their representations, a Solid server may lose formatting details and comments, even when you store a Turtle file and then retrieve it again as Turtle.
The container description used by Solid is the LDP BasicContainer, so direct children are listed with ldp:contains triples.
All operations on ACL documents act the same way as on other RDF resources, except that they always require Control access over the resource to which the ACL applies, regardless of whether it's a reading, creating, appending, modifying, or deleting operation. There is now way to know if a resource is an ACL document, and if it is, there is no way to know to which resource it applies, except that the resource to which it applies will link back to it with a Link response header (relation: 'acl').
HEAD and GET on non-ACL documents both require Read access.
There are three ways to create a new resource: POST, PUT, and PATCH. POST adds a new resource inside a container, and although you can specify a Slug header to influence the resulting location, the server decides the URL at which the resource will be created and report it to you in a Location response header. For PUT and PATCH, the client can choose the location.
Empty containers can only be created with PUT or PATCH, because it requires forcing the URL to end in a slash. Note that all containers also act as RDF sources, so you can add triples to them and edit those (as long as you don't try to directly add or edit the container's containment triples).
In all three cases, the server will create any ancestor containers on the path from the domain root to the URL of the request if they don't exist yet. If a container needs to be created then it requires Write access on that container and Append or Write access on its parent.
A POST request requires Append or Write access on the container in which the resource is being created. A PUT does too, but additionally it requires Write access to the URL of the newly created resource itself. It's not possible to create an empty container, other than by creating a dummy resource inside it (and then deleting that again).
PUT requests should always have either an If-None-Match: * header (to avoid overwriting anything) or an If-Match: "[etag]" header, to make sure the request will not write over any changes that may have happened since the app last read the resource. When using PUT to create or update RDF sources, if what you want is "on conflict do update", consider using PATCH instead.
Non-container RDF sources (including ACL documents) can be edited and created using a PATCH request with a (restricted) sparql-update body. The sparql-update body can include INSERT and DELETE instructions. If one of the DELETE instructions fails, the request fails as a whole and is not executed. PATCH requests that contain only INSERTs require either Append or Write. PATCH requests that also contain DELETEs require Write + Read. Like for POST and PUT, PATCH will cause all ancestor containers to be created if they are missing, and the whole operation will fail if permissions are insufficient.
A DELETE of a resource requires Write access to the resource and Write to the container it's in. A container can only be deleted if it's empty.
Servers are in theory allowed to be eventually consistent, although this is not recommended. This means that when you do a PUT and then immediately do a GET, if the PUT went to a master node in a master-slave cluster, and its effect did not propagate yet to the slave that handles your GET request, the GET may be completed based on outdated information. Likewise, in a master-master setup, two master nodes may each accept a write operation and report success (200 OK response), but the cluster may then later roll back one of them if the two writes are not reconcilable. It's probably easier for both the app developer and the end user to work with a solid pod that is not only solid spec compliant, but also strictly consistent.
Authorization type (acl:Authorization)
An ACL document can grant access to various patterns of requests, using a list of additive Authorizations. Each authorization needs to have a URI that dereferences to a fragment within the ACL document itself, and have RDF type acl:Authorization.
Authorization target (acl:accessTo, acl:default)
An ACL document on a non-container affects only that resource, through the acl:accessTo predicate. An ACL document on a container affects that container through the acl:accessTo predicate, but it also affects all its descendants through the acl:default predicate, as long as they don't have their own ACL document. So conversely, if a resource doesn't have its own ACL document with acl:accessTo authorizations, then the acl:default authorizations from the ACL document of the container it's in applies, unless that doesn't exist either, etcetera, until you reach an ancestor container that does have an ACL document. Only one ACL document ever applies, and rules from different ACL documents are never mixed together.
If the ACL document that applies directly doesn't have any (matching) acl:accessTo authorizations, or if that ACL document doesn't exist and the first ancestor ACL document that does exist doesn't have any (matching) acl:default authorizations, then all access to that resource is denied ("deny then allow" approach).
Even for an authorization that does apply to the resource in question, there are still three dimensions in which it has to match before it can actually grant any access, as the following three sections will explain.
The first is modes (Read, Append, Write, Control). The previous section mentions which access modes are required for each operation. Note that if you include Write then there is no additional effect in also including Append.
The second is agent. There are four ways to add agents:
- Using acl:agent with a specific WebId that should be granted access
- Using acl:agentGroup with a vcard:Group, whose members should be granted access
- Using acl:agentClass with acl:AuthenticatedAgent, meaning any correctly authenticated WebId
- Using acl:agentClass with foaf:Agent, meaning anyone (public access, no authentication needed)
The third dimension is application; for authorizations that include public access in the agent dimension, all origins are allowed in the application dimension. If public access is not granted in the agent dimension, then only the storage server's own (same) origin is allowed by default in the application dimension. All other origins are only allowed if they are listed on the Authorization, using the acl:origin predicate.
Controlling which apps can access which data
A user can use an ACL editor or a launcher app to edit the ACL documents on their pod. Usually the user will want to have full access themselves to all subtrees of containers on the pod, but only at the same origin and at apps the user really trusts (for instance a launcher app or a command-line tool for power-users).
Some public access will also often be allowed, for instance POSTing to an inbox, or reading documents from a public area on the pod. Public access does not require the user to identify themselves, and neither does it require the application to do so. When the pod owner themselves access data, they generally have Control access, which allows them to edit the access control lists, and grant the apps they want to use access. If other users use the same apps as you do, then you can also share resources with them inside those apps.
But note that when Alice gives Bob acces to a resource on her pod, then Alice decides which apps Bob can use to access that resource. Bob does not get the ability to add or remove any apps from that list. This is a known issue for which Solid's app authorization panel is currently searching a solution.
On each resource, you'll find an Updates-Via response header, pointing to a WebSocket server. Connect, and send auth [bearer_token], then sub [url]. You'll get ack [url] back, and then pub [url] each time the resource changes. A change can be a creation, update, or deletion. If you subscribe to a container, you'll only see creation and deletion of the container itself, and changes to its list of ldp:contains triples, not any changes to the resources it contains.
Getting your app listed
To get your app under the attention of Solid users, try to get it onto https://solidproject.org/use-solid/apps by sending a pull request to https://github.com/solid/solidproject.org/blob/staging/pages/use-solid/apps.md. There is also a list of available apps in inrupt's launcher exploration, although this is still experimental. A manifest format in which apps can self-announce which permissions they need on a user's pod is probably coming soon, and then any launcher app can retrieve that machine-readable manifest - either from the app's origin or from some Solid app registry. Until then, we can also simply update the various launchers manually to add our apps to their lists.
If your app uses data domains for which the data conventions have not yet been added to the client-client spec then you can try to claim the right of first arrival, and propose a format to use for that data domain. Once the client-client spec has been created, I'll update this paragraph to link to it.
In mid-2019, and again at the end of 2019, the Solid spec was updated. The version described here is the December 2019 one. It differs from the mid-2019 one in the following ways:
- Globbing, sparql-on-get, and acl:trustedApp were removed.
- Servers are in theory allowed to be eventually consistent, although this is not recommended.
- WebSockets-pubsub clients SHOULD now send AUTH, but servers should not require it yet.
- Apps SHOULD support both webid-oidc flows, but IDPs and storage providers SHOULD still support the old one.
- IDPs are no longer required at the spec level to offer a WebId-TLS bridge (that's now a consideration that's between them and what their own users want).
- The auth command in WebSockets-pubsub will become mandatory.
- Storage servers should switch to the new WebId-OIDC bearer token format.
- IDPs should switch to new WebId-OIDC flow.
- Once all storage servers and IDPs had a chance to switch, apps should stop supporting the old format.