The PermawebOS is here: 10 new Releases
Today we are shipping 10 new components and massive improvements.
In the comin | Hanami
The PermawebOS is here: 10 new Releases
Today we are shipping 10 new components and massive improvements.
In the coming era, nothing is certain. The value of specific software products has dropped to virtually zero.
We think that the only thing that will continue to hold value in this new world is protocols: the mutually agreed means of getting things done -- where each party cannot simply sling out a custom version upon request. The larger the group that needs to agree, the harder it is to change.
We have been building AO exactly for this new age: a universal computation protocol that we think will give rise to a decentralized, shared supercomputer. It is a new type of thing: a protocol built to house and execute other protocols, while sharing infrastructure freely and re-usably between all of them, and allowing each to easily gain the properties of trustworthiness, permanence and decentralization.
Alongside that computer and its universe of shared protocols is Arweave: a decentralized, open data lake with over 25 billion pieces of content, each of which can be re-used and incorporated trivially by any new app or experience.
In totality, we call this system the PermawebOS. It has three core layers:
Arweave: the data lake, now with fully trustless reads and writes that you can join and get paid for.
AO: the backend of all compute in the network. Shared protocol implementations, executable anywhere.
UIs: a single meta-app across all platforms -- web, browser extension, mobile -- that allows you to conjure any experience you want with the data on the permaweb.
The PermawebOS is an enormous project. We built its first component -- Arweave -- almost 10 years ago now. Today we are shipping 10 new components and massive improvements.
1. PermawebOS on Android
A trustworthy PermawebOS node in every pocket.
The easiest way to run a secure AO node is here, and it's as simple as installing an Android app. The PermawebOS mobile node runs HyperBEAM on the phone and binds it to evidence about the device: the app build and the key that signed it, a hardware-backed node key, Verified Boot and bootloader state, and the security patch level, all of which a zone can check against its own policy before admitting the node or sending it work.
Phones are the right hardware class for decentralizing AO -- there are a lot more Android phones than enclave servers, and you probably have a spare.
Today, phones can join the AO network as bundlers for up.arweave.net, execute processes, and run as the backend for apps you build.
Try it: install the PermawebOS Android app (download .apk) on a spare, up-to-date Android phone with a locked bootloader, configure it in the new extension (download .crx).
If Chrome won't install the .crx, use the unpacked extension ZIP instead: extract it, open chrome://extensions, turn on Developer mode, click Load unpacked and select the extracted chrome folder.
2. Agents as AO processes
Your new PermawebOS mobile node can host agents with first-class access to the permaweb, a full Linux VM, and 100+ HyperBEAM devices: a complete development environment out of the box.
Build an app with an agent today and you still need someone's cloud to run it. Build it on PermawebOS and it runs on hardware you own and can prove is trustworthy, with the result kept permanently on Arweave.
[email protected] is a stateless tool loop. Hand it a set of tools and a model and it calls [email protected], runs whatever tool calls come back, and repeats until the job is done. Put that loop inside an AO process and the process is the agent. Each scheduled message is a turn: the process builds a prompt from its own state, resolves [email protected], and stages whatever the agent decided to send. Mailbox, thoughts, tool policy and model settings all live in process state on the message log, which means no second runtime somewhere is keeping a private copy of the conversation. Agents only talk to the world through explicit messages, the same way every other process does.
Try it: the agent stack ships in Ouroboros, which runs as a HyperBEAM device bundle on any node (see #10).
3. Direct LLM inference as AO devices
Direct LLM inference is available as devices in AO: Apus, local models, OpenAI, Anthropic, OpenRouter and Cerebras.
[email protected] is an OpenAI-compatible bridge with native Anthropic support. Out of the box it talks to OpenAI, OpenRouter, and local models through Ollama or llama-swap, and its custom provider covers any other OpenAI-compatible API, Cerebras included. Apus runs its own [email protected] on its AO-powered GPU network, and you don't need a key to call it. On PermawebOS for Android, a local inference device runs the model on the phone itself -- every inference call is an ordinary AO message, so the prompt and the answer land on the hashpath along with everything else the node computed.
Run through the PermawebOS mobile node setup with the extension, and connect your own inference providers -- get the extension here.
Read more: The Permaweb as a Substrate for AI.
4. Full Linux and VM execution inside AO
Full Linux and VM execution inside AO: QEMU, Docker, and even on Android phones with [email protected].
Plenty of agent work needs a real computer: a compiler, a package manager, and a shell that can be used for tool calls. [email protected] gives each agent its own persistent virtual machine, [email protected] gives it a persistent container, and on Android [email protected] gives it a persistent Ubuntu userland on the phone it is running on. This kind of power on mobile devices would usually come at the cost of using somebody else's cloud. With PermawebOS, it's possible on your phone -- and your phone runs the public backends for apps you build, too.
Try it: enable [email protected] or [email protected] in an Ouroboros node config; [email protected] runs on the PermawebOS mobile node.
5. AO-Core hashpaths v2
Trace not just how any value in any computation was created, but also every other branch in the tree that would have led you to the same conclusion.
The first hashpaths were a rolling hash of the previous hashpath and the request, which was enough to show how a value was made. A v2 hashpath is an executable ao:// URI. It records the request, the inputs the device declares it reads, where each of those inputs came from, and the result. Since the record is cut down to the inputs that mattered, any other route that reduces to the same inputs reaches the same result, and a node can show you that it does. Loading a hashpath re-runs the computation from scratch; verifying it is a separate step that checks the claimed result against what the re-run produced.
Try it: every response from a HyperBEAM node already carries one. curl -sD - -o /dev/null 'https://arweave.net/[email protected]/set&a=1' | grep -i hashpath prints the hashpath of that tiny computation, and the node signs over it.
6. HyperBEAM kernel v0.11
HyperBEAM v0.11 brings over 700 fixes across the stack, leading to performance improvements ranging from 2x to 11x for some activities and the finalization of the model for AO-Core 1.0.
The last stop before AO-Core 1.0. Only message extensions now stand between v0.11 and 1.0. Most past bugs traced back to open questions about what data is and what a computation may depend on. Those now have precise answers, so the fixes are mechanical, and node storage moves to LMDB 1.0, the database's first major upgrade in years.
A permanent library of every computation. Almost every computation ever run has been done in private and thrown away, so machines across the internet repeat the same work endlessly. v0.11 reduces each computation to exactly the inputs its device reads before running it (a step called varying, after HTTP's Vary header), so a result is identified by what it depends on and nothing else: a transfer of three tokens from a balance of seven is the same transfer on every node. Results are kept as compact records that reference their inputs and output rather than copying them, and nodes publish them to Arweave as LMDB databases that other nodes read in place. A node checks whether work has been done before doing it, a new node starts warm by importing and verifying what its peers have worked out, and every new computation makes the next one cheaper for everybody.
No more cache poisoning. Caching is where computers most often get things wrong. In earlier versions, a device that modified its input could return a result still carrying the original's signatures, and the cache would file it under the original's ID. v0.11 normalizes results before storing them and files each one under exactly the inputs it read, so anyone can check an entry against its hashpath.
Types, and nodes that document themselves. Devices now declare their inputs in ordinary Erlang type specs. HyperBEAM resolves links and checks types before the device runs, so device code shrinks to its logic. The same types let every node serve docs at /[email protected] for exactly the devices it has loaded, with runnable requests, so an agent given nothing but an address can find its way around. Arweave.net and ao.arweave.net are already on it.
The PermawebOS mobile node is built on v0.11. Get the code for the kernel here.
7. Trustworthy TLS to the node
Today the padlock in your browser only proves which web address you reached, not who is actually running the machine behind it. Tying the certificate to the node's Arweave wallet closes that gap: the certificate behind the padlock now holds the same key that signs the node's results, stakes on the network and gets paid for its work. You know exactly whose word you are trusting, with no extra software to install.
On the permaweb, you can now verify that you are talking to a specific Arweave wallet -- starting from the browser's padlock icon.
With [email protected], a HyperBEAM node terminates HTTPS itself and uses its Arweave wallet key as its TLS key. It will refuse to load any certificate whose public key is not that wallet key, and it never falls back to plaintext. Operators can take a signing request from GET /[email protected]/csr to any certificate authority, or let the built-in ACME client issue and renew the certificate for them (HTTP-01, or DNS-01 through [email protected]). The certificate's public key is the node's wallet key, but browsers present it differently from Arweave: Chrome shows a hex SHA-256 fingerprint of the whole key, exponent included, while an Arweave address is the base64url SHA-256 of the modulus alone. The two won't match by eye. Both come from the same key, though, so the padlock always leads back to the node's wallet.
8. Trustworthy real-time indexing and querying
Over a terabyte of Arweave's transaction index now lives on the weave itself, packed at about 9.5 bytes a row. Any browser can leaf through it a few chunks at a time, straight from Arweave, without anyone downloading the whole index or trusting an endpoint to serve it.
Live and up-to-date trustless indexing on the largest PermawebOS node, arweave.net. Historical information queryable client-side, too.
Arweave.net's GraphQL answers from HyperBEAM's realtime query layer. One query matches across the history published on Arweave as ArLMDB snapshots, confirmed transactions the node has indexed itself, and uploads that are paid for and verified but not yet in a block. Every result carries a hashpath and an attestation from the node, so nothing is shown "optimistically" because a bundler said so. The history half is a compact index stored permanently in a format a browser can walk a few chunks at a time. That means your own browser can run the historical part of a query without trusting anybody's endpoint.
Try it: arweave.net/graphql for live queries, and GraphQL running in your browser straight against the weave.
Read more: Arweave is a realtime database and Self-sovereign GraphQL in every browser.
9. Bundling pools
Provide bundles from your phone, making up.arweave.net decentralized, and helping users upload to the permaweb.
Most permaweb apps already upload through up.arweave.net. Behind that address is now a pool of bundler nodes that anyone with a phone can join. Choose the bundler capability when you configure your mobile node, join the zone from the Bundle Data page on ao.arweave.net, and keep some AR in the node's wallet to settle data on Arweave. Your node accepts uploads and bundles them via the up.arweave.net router. Because indexing now reads straight from miner mempools (see #8), an upload through a small operator's phone is discoverable just as fast as one through a big provider.
Try it: this is open to volunteers today -- paste your node's tunnelled address in the AO Discord to join the zone.
Read more: The Decentralized Bundler Network is Coming and PermawebOS LapEE Alpha.
10. The full-stack PermawebOS: browser extension and Ouroboros
Everything in this release ties together in two parts of the full-stack PermawebOS that ship today.
Browser extension: pair your mobile node to talk to your sovereign agent. The PermawebOS extension launched in August as a wallet and an AO node at the browser boundary. Its chat tab is now open to everyone. Connect your phone and the extension creates an agent on it, lets you pick a model (your ChatGPT subscription, or a key from another provider), and drops you into a conversation. The agent lives on your phone and acts under your keys, and the whole conversation is AO messages you can audit.
Ouroboros: sovereign agents for teams and communities. Ouroboros packages the agent devices from #2 and #3 with a web UI for long-lived, collaborative agents. A space is itself an AO process that holds its members, child spaces and admission rules, and both people and agents can be members. Private messages travel as WeaveMail, encrypted to the recipient's own key. A team runs its spaces on nodes it controls, so the agents and everything they remember stay with the people who use them.
Try it: install the PermawebOS extension for Chrome and open the chat tab.
Ouroboros lives on the permaweb at arweave://ouroboros.
Read more: PermawebOS: An AO Node in Every Browser.
...and any app you can dream up
The point of each of these is not that they are each a product. Each is a fully re-usable cog that can be tweaked and morphed into anything the user wants. All share the same infrastructure, data, and protocols.
For example, Bugcraft: deploy intelligence swarms to solve tasks for you (early alpha).
Bugcraft's first swarm hunts bugs. A lead agent sends searchers through a codebase, gives each issue its own agent to write the patch, and puts every line in front of a human before anything merges. The code lives on the permaweb at arweave://bugcraft, where anyone can fork it and point a swarm at their own kind of task.
Another example: we are using Ouroboros to build Ouroboros, recursively, and have deployed it for anyone to collaboratively build the AO ecosystem. Soon, you'll be able to join the public Ouroboros space on AO to take part.
The permaweb is here. To build it, all you have to do is ask.
Download the PermawebOS extension to pair a mobile node and get started with your personal agent.
Read this on the Permaweb:
(Historical earnings: For fiscal 2026FY (period ended 2026-09-03), Micron reported basic EPS of 75.38, diluted EPS of 74.33, net income of USD 84.969 billion, and revenue of USD 133.188 billion.
Consensus expectations: For the next upcoming period (2026Q4, event date 2026-09-30), consensus expects EPS of 32.5625; no revenue estimate is available.)