ElizaOS Hits a New Low as the Eliza Framework Keeps Shipping

ELIZAOS, the token that grew out of crypto’s most famous AI agent experiment, touched a fresh all-time low of $0.0003419 on August 5, 2026, according to CoinGecko. That is not, on its own, surprising; the token has been grinding lower for a year and a half. What is more interesting is what is happening at the same time on the other side of the project: the open-source Eliza framework that ELIZAOS is named after just closed out one of its busiest development stretches yet, and a publicly traded company is weeks away from shipping a commercial product built on top of it.

That split, a token near zero and a codebase still shipping weekly releases, is the actual story of Eliza in the second half of 2026. This piece sets aside the well-worn rise-and-fall narrative (the Marc Andreessen tweet, the lawsuit, the migration that multiplied the token supply tenfold) to look at what the framework actually is under the hood: how its plugin architecture works, who is building on it now, where the real security risks sit, and whether “the framework outlives the token” is a genuine thesis or just a comforting story for a project with a court date.

What the Eliza Framework Actually Is (and Isn’t)

Eliza is easy to describe badly, because the name has been attached to several different things since October 2024. The Eliza framework itself is a free, MIT-licensed TypeScript codebase, maintained on GitHub under the elizaOS organization, for building AI agents that hold a persistent character, connect to chat platforms and blockchains, and take actions on a user’s behalf. The repository describes itself simply as “your agentic operating system,” which is either an accurate mission statement or a hint of scope creep, depending on how the next few sections land.

Separate from the framework is Eliza Labs Inc, the company Shaw Walters founded to build it, and the Eliza Foundation, the nonprofit entity that now stewards the open-source project and signs partnership deals like the one with Secure Blockchain covered below. Separate again is the AI16Z DAO, the community of original token holders, and the two tokens themselves: legacy AI16Z, which still trades in thin volume on Solana decentralized exchanges, and ELIZAOS, the migrated token that has been the primary one since late 2025. A reader can reasonably trust the framework’s engineering, distrust the DAO’s governance, and hold no opinion at all on either token; these are not the same bet, and treating them as one is exactly the confusion this piece tries to untangle.

A $75,000 Bet That Became a Cautionary Tale

The origin story has been told in detail elsewhere, so the short version: developer Shaw Walters launched ai16z in October 2024 on Solana through the daos.fun launchpad, targeting a raise of around $75,000 to fund an AI-managed trading experiment. Two agent personas fronted the project, Marc AIndreessen, styled after venture capitalist Marc Andreessen, and Degen Spartan AI, a more aggressive trading persona. On October 27, 2024, the real Marc Andreessen tweeted “GAUNTLET THROWN” in apparent approval, and the token’s market cap spiked to roughly $96.6 million within hours, briefly overwhelming the launchpad’s site, according to The Block’s contemporaneous coverage.

Within days, the crypto publication Protos reported that the supposedly autonomous trading agent was, in practice, run by humans who approved trades manually, a gap between marketing and reality that later became the core of a federal lawsuit. The token kept climbing anyway, peaking near $2.47 in the first days of January 2025 with a fully diluted valuation above $2.6 billion. Later that month, Andreessen Horowitz asked the project to distance itself from the AI16Z name, and the team rebranded the DAO and its software to elizaOS on January 28, 2025. A token migration in October and November 2025 then expanded the supply roughly tenfold, from about 1.1 billion to a target near 11 billion units, a detail that would later show up in a lawsuit as an insider-dilution allegation.

That lawsuit, Pikabea v. Walters, is still working its way through a New York federal court and gets its own section further down. What matters for the rest of this piece is what kept running underneath all of it, largely ignored by anyone just watching the price chart: the open-source codebase.

Inside the Architecture: Actions, Providers and a Plugin for Everything

Strip away the token and the marketing, and Eliza is a fairly conventional idea executed with unusual thoroughness. An agent starts as a character definition, a structured file describing a name, a biography, a body of example dialogue, a knowledge base and a style guide, which the runtime uses to keep the agent’s personality consistent across every platform it touches. That runtime is model-agnostic: it can call OpenAI, Anthropic, Google’s Gemini, xAI’s Grok, open-weight Llama models, or the project’s own locally hosted Eliza-1 models, which range from 2 billion to 27 billion parameters and are built on Google’s Gemma-4 architecture.

The organizing idea behind the current architecture, according to the project’s own documentation, is that almost everything is a plugin. Database adapters, blockchain clients, chat connectors and the smaller building blocks that make up an individual skill all implement the same Plugin interface and register through the same lifecycle: the runtime validates that a plugin has a unique name, runs its optional setup function against the current configuration, then registers its components (adapters, actions, evaluators, providers, models, HTTP routes and event handlers) in a fixed order so dependencies load before the things that depend on them.

Four primitives do most of the conceptual work, and they are worth knowing if you are trying to evaluate any agent built on the framework, not just Eliza’s own reference agents.

Worlds, Rooms and a Fast-Moving Release Cycle

Above the plugin layer, Eliza organizes context using a small set of nested concepts: a World is roughly a server or workspace, a Room is a channel or a direct-message thread inside that world, and an Entity is any user or agent participating in it. The point is that an agent can hold a separate, coherent memory per room while still being able to signal or coordinate with other agents across rooms, which is what lets a project run several specialized agents (a support agent, a trading agent, a moderation agent) that share an underlying identity without stepping on each other’s context or leaking one user’s private conversation into another room’s memory.

That messaging layer has been in active flux for most of 2026. The project has been working through a shift toward a more explicit channel-based architecture for how messages and rooms map onto each other, part of a broader v2 effort that was previewed at Jupiter’s Catstanbul event back in early 2025 and has been shipping in beta ever since rather than landing as one finished release. The pace of that shipping is genuinely fast: in one representative week in late June 2026, the repository’s release history shows three beta tags landing back to back, v2.0.3-beta.5 and beta.6 on the same day and beta.7 two days later, covering everything from cloud sign-in fixes to local text-to-speech decoder support to mobile touch-target sizing.

None of that is glamorous. It is also exactly the kind of unglamorous, weekly-cadence maintenance that a token chart cannot show you, and taken together with the issue tracker, it is the strongest evidence that the engineering team did not stop working when the price did.

Local Models and a Quiet Bet on Becoming an Operating System

Two details point to where the project wants to go next, beyond being a library developers import into a script. The first is Eliza-1, the framework’s own family of locally hosted models mentioned above. Running inference locally instead of calling a third-party API cuts latency and cost for high-frequency agent actions and removes a dependency on any single model provider’s uptime or policies, which matters for a framework whose entire pitch is that agents should be able to act continuously and autonomously rather than pause every time an external API has a bad day.

The second, more ambitious signal is a separate repository, elizaOS/os, described as the home for an Android Open Source Project fork, a Debian-based live Linux distribution, installers and image-building tools. The framework itself, its mobile apps, native device bridges and local inference stay in the main eliza repository; the os repository consumes those published components to build actual bootable system images. In plain terms, the team is laying groundwork for agents that run at the operating-system level on a phone or a dedicated device, not just inside a chat window or a browser tab.

It is worth being direct about where that stands today: the os repository has essentially no GitHub traction yet, a single star at last check. This is a stated direction, not a shipped product, and it is fair to treat it as a roadmap item rather than evidence of anything already working at scale.

The Plugin Registry Economy: Who Is Actually Building

The clearest non-price evidence of whether a framework is actually being used is whether outside teams bother to maintain their own tooling around it. On that measure, Eliza’s ecosystem looks healthier than its token. Beyond the official plugin registry that syncs through the project’s own documentation site, at least two independent infrastructure teams run their own registry forks: Automata Network, an onchain attestation protocol, and Nethermind, one of the larger Ethereum client and infrastructure teams in the industry.

Automata’s involvement is more than a passive fork. The team maintains a proof-of-agent plugin that lets an Eliza agent’s execution be verified through Intel’s DCAP remote attestation process, essentially a cryptographic receipt that an agent ran inside a trusted, tamper-evident enclave rather than an environment an operator could quietly alter after the fact. That plugin sits alongside a broader push toward verifiable agent execution across the sector, the same motivation behind EigenLayer’s EigenCloud push into verifiable AI and, for workloads where trusted hardware isn’t an option, optimistic verification approaches like opML.

The plugin catalog itself spans core infrastructure (message handling, SQL storage, retrieval-augmented knowledge bases), blockchain connectivity across more than thirty EVM networks plus Solana, platform clients for Discord, Telegram, X and Farcaster, and interchangeable large language model providers. None of that requires anyone to hold ELIZAOS. That is precisely the point critics of the “framework outlives the token” framing make: the plugin economy can keep compounding in a way that never flows back to token holders at all, which is good for developers and largely irrelevant to anyone still holding the token for its price.

What Happens When an Agent’s Memory Can Lie to It

The framework’s biggest technical liability is not a bug so much as a structural property of how persistent-memory agents work at all. In March 2025, researchers from Princeton University and the Sentient Foundation published a paper, titled “Real AI Agents with Fake Memories,” that used Eliza as its primary case study. Their framing was blunt: “AI agents integrated with Web3 offer autonomy and openness but raise security concerns as they interact with financial protocols and immutable smart contracts,” the authors wrote in the paper’s abstract.

The specific attack they demonstrated, memory injection, is more dangerous than ordinary prompt injection because it persists. A prompt injection tries to trick a model within a single exchange; a memory injection gets an attacker’s fabricated message stored inside an agent’s long-term memory as if it were a verified fact. Later, potentially on a completely different platform than where the injection happened, the agent retrieves that memory as legitimate context and can be walked into authorizing a transfer or violating a protocol rule it would never have agreed to if the fake memory had been flagged as untrusted input. The researchers, Atharv Singh Patlan, Peiyao Sheng, S. Ashwin Hebbar, Prateek Mittal and Pramod Viswanath, built a benchmark called CrAIBench specifically to measure how exposed different agent frameworks are to this class of attack, and reported that simple prompt-level guardrails were not sufficient defense on their own.

Eliza is not the only framework this affects, and the failure mode is not unique to any single project; researchers and security firms have documented similar memory and prompt manipulation across the sector, including cases where a hidden instruction embedded in something as mundane as a social media reply triggered an unauthorized six-figure transfer from an unrelated agent’s wallet. What made the Eliza-specific case notable enough for Decrypt to describe an agent as being “gaslit” into losing money is simply that Eliza was, by GitHub activity, the most widely deployed framework the researchers could study at the time.

The ecosystem’s response has split into two camps. One is verifiable execution: prove cryptographically that an agent ran the code it claims to have run, through TEE attestation like Automata’s plugin above or restaking-secured verification layers. The other, older camp is just good custody hygiene, treating an autonomous agent’s wallet the same way a treasury team would treat any hot wallet: spending limits, address allowlists, time locks and a human-in-the-loop check above a certain transaction size, the same principles covered in HOGE Wire’s multisig security playbook. Agents that bridge assets across chains to execute a strategy inherit a third risk layer on top of both: the bridge itself, a category HOGE Wire has tracked through 2026’s ongoing wave of bridge exploits, none of them agent-specific but all of them relevant to any agent moving funds cross-chain on a user’s behalf.

Eliza vs. the Rest of the Agent-Framework Field

Zoom out from crypto and Eliza is a small player. LangChain, the dominant general-purpose framework for building LLM applications, carries well over 100,000 GitHub stars, several times Eliza’s count, and AutoGPT, the project that popularized the term “autonomous agent” back in 2023, sits in a similar range. Neither was built with a crypto wallet or a token launch in mind; both get bolted onto Web3 use cases after the fact rather than starting there, which shows in how much extra plumbing a developer has to write to give either one native chain awareness.

Inside the narrower crypto-native segment, the comparison looks different. Virtuals Protocol’s GAME framework takes an almost opposite philosophy to Eliza: no-code agent creation tightly integrated with its own Base-based launchpad and the VIRTUAL token, currently trading around $0.56 with a market cap near $370 million according to CoinGecko, optimized for getting a monetizable agent live quickly rather than for developer flexibility. Rig, a Rust-based framework, trades Eliza’s TypeScript accessibility for raw performance and has a smaller, more specialized following. Smaller entrants like ZerePy and ARC occupy narrower niches still, generally built by single teams for their own product rather than as general-purpose infrastructure.

Raw star counts flatter the general-purpose frameworks, since they serve a developer population many times the size of crypto’s. A fairer read is that among agents needing a native wallet, multi-chain actions and a persistent cross-platform character out of the box, Eliza remains the default open-source choice, which is exactly why it, rather than a general-purpose framework, ended up as the shared substrate under Clawville, Agentic SME and the rest of the deployments below.

The Test Case: Clawville, Agentic SME and a Mid-August Deadline

The strongest argument for the framework’s health is what other teams choose to build with it. Clawville, an AI-native open-world MMORPG built inside the Milady ecosystem, runs player-deployed agent companions on the elizaOS runtime, letting a player either control a character directly or let an agent autonomously grind quests, gain experience and evolve inside the game world using the same persistent-memory system described above, according to Cointelegraph’s coverage of the launch.

Funding followed developers rather than the other way around. At Jupiter’s Catstanbul event, the Solana-based exchange partnered with elizaOS to launch a $10 million Magic Fund aimed at backing open-source agent developers, separate from and unaffected by whatever the token happened to be doing that week, per The Block’s reporting from the event.

The more consequential test is Agentic SME. On April 10, 2026, TSXV-listed Secure Blockchain Development Corp completed an all-stock acquisition of Agentic Solutions Ltd, appointing Agentic co-founder Sebastian Quinn-Watson, also, awkwardly, a named defendant in the lawsuit below, as a director. Five weeks later, on May 15, Secure Blockchain and the Eliza Foundation finalized a development agreement to co-build Agentic SME, a white-label product that automates sales, admin, finance and customer-support work for small and mid-sized businesses on top of the open-source framework, according to the companies’ own announcement. The build plan called for an initial version within thirty days and a final deliverable within three months, putting the target launch window in mid-August 2026, a matter of days from when the token was busy setting a new price floor.

That timing is the real story, not a coincidence to gloss over. A publicly traded company shipping a commercial product has to satisfy ordinary corporate and customer liability standards, not crypto-native risk tolerance. If Agentic SME actually ships and holds up in production, it is a materially stronger validation of the underlying framework than any amount of Discord activity, precisely because Secure Blockchain has no reason to sugarcoat a broken dependency the way a token-aligned community might.

The Token Side: A Record Low the Same Week as the Biggest Launch Yet

None of the framework’s technical health has translated into the token. ELIZAOS was trading around $0.00034 as of this writing, just above the fresh all-time low of $0.0003419 it set on August 5, 2026, for a market cap near $2.56 million on roughly 7.48 billion circulating tokens, per CoinGecko. The token’s high-water mark was never really a price peak at all: its highest market capitalization, just under $10.9 million, came on November 6, 2025, migration launch day, and has been sliding more or less continuously since.

The original AI16Z token, the roughly 1.1 billion units that never migrated, trades even thinner: about $0.0003 apiece, a market cap near $327,000, and daily volume that barely clears five figures, according to CoinGecko. It set its own all-time low, $0.0001919, on July 24, 2026, less than two weeks before ELIZAOS did the same.

Put the two halves of this story side by side and “the framework outlives the token” starts to look less like a prediction and more like a description of something that has already happened. The open-source project and the token that shares its name are, at this point, two separate assets with two separate trajectories, and nothing about Agentic SME’s success or failure this month is likely to move either one back toward the other.

The Lawsuit Is Still Stuck on Step One

The legal cloud over the project’s founders has not lifted. Gorka Pikabea, a Spanish investor represented by Burwick Law, filed a class-action complaint against Eliza Labs Inc, Shaw Walters, Sebastian Quinn-Watson and the AI16Z DAO in the Southern District of New York on April 20 and 21, 2026, docketed as case 1:26-cv-03238. The complaint alleges violations of New York and California consumer-protection law, negligent misrepresentation and unjust enrichment, centered on the claim that the project marketed a human-operated trading operation as a fully autonomous AI fund and diluted existing holders through a migration that expanded token supply roughly tenfold.

As of the most recent public hearing, in mid-May 2026, the case had not moved past serving the defendants. Plaintiffs’ counsel told the court they had made a single attempt to serve Walters at a San Francisco address. Judge Jed S. Rakoff was unimpressed, telling the parties on the record, “I am not going to move the case forward on this basis: 1 attempt?” according to court reporting from Inner City Press. He ordered a written accounting of service efforts and pushed the next hearing back two weeks. Quinn-Watson, based in Australia, requires service through the Hague Convention process, which court watchers estimate could take months on its own.

No defendant has responded on the merits, and nothing in the public record as of early August 2026 shows the case has progressed beyond that service dispute. The allegations remain just that, unproven in court. The one detail worth sitting with is Quinn-Watson’s dual role: ten days before Pikabea’s lawsuit named him as a defendant accused of overstating ai16z’s autonomy, his company was acquired by Secure Blockchain specifically so he could build commercially on the same open-source framework at the center of the claims against him.

Where U.S. Regulators Currently Stand

For American readers, the relevant backdrop is the joint interpretive release the SEC and CFTC issued in March 2026, formally published in the Federal Register on March 23, according to Forbes’ coverage of the release. It sorts crypto assets into five categories, digital commodities, digital collectibles, digital tools, stablecoins and tokenized securities, and turns the securities question on whether a project promised “essential managerial efforts” that would lead purchasers to expect profits, versus vague claims lacking any concrete milestones or funding plan.

The release does not mention AI agents, autonomous software or agent-driven trading as their own category at all, which leaves the exact question at the heart of the Pikabea complaint, whether marketing an agent as autonomous when it is not amounts to the kind of managerial promise that triggers securities treatment, unresolved by regulation and left to this one lawsuit to test in court instead. Readers who want the fuller picture of how aggressively the SEC has actually pursued crypto cases so far this year may find HOGE Wire’s report card on the agency’s own enforcement record useful context.

Is It Still Worth Building on Eliza in Mid-2026?

Strip out the token and the litigation, and the case for the framework itself is genuinely strong: 408 open issues and 83 open pull requests being actively worked, over 11,000 commits on the development branch, a weekly beta release cadence, multiple independent teams maintaining their own plugin registries, and a publicly traded company betting a commercial product’s launch date on it.

The case against building on it is not the code, it is what surrounds the code: a founder and a co-founder named as federal defendants, a token with no institutional credibility left to lose, and a security research literature that keeps finding the same memory and prompt manipulation weaknesses common to every framework in this category, not just this one. Those are separable risks, and a developer shipping a plugin or an internal tool on Eliza is exposed to the code-level risk, not the litigation or the token, unless they choose to tie their own product economically to ELIZAOS specifically.

Shaw Walters has been fairly consistent about where he thinks this goes. Writing in Forbes in January 2026, he argued that “AI agents won’t go mainstream because they’re intelligent,” predicting instead that “they’ll go mainstream because they’re quirky, personal, and sweep through culture the way memes do,” in his own column. Whether that turns out to be right or wrong, it is a bet about the framework and the agents built on it, not about a token chart, and 2026 has been the year that distinction stopped being theoretical and started being the entire story. The more interesting question going into next year isn’t whether ai16z was a fraud; a court will eventually decide that. It’s whether an open-source agent framework can keep compounding developer trust while its own namesake token becomes a cautionary tale on a chart, a split that AI-crypto as a sector is going to keep running into as more frameworks spin up speculative tokens around genuinely useful open-source infrastructure.

Frequently Asked Questions

What is the Eliza framework?

Eliza is a free, open-source TypeScript framework for building AI agents that can hold a persistent character, connect to platforms like Discord, Telegram and X, and take actions such as sending messages or executing blockchain transactions. It is maintained under the elizaOS organization on GitHub and licensed under MIT, meaning anyone can use, modify or self-host it without paying a fee.

Is ElizaOS the same thing as ai16z?

They are related but not identical. Ai16z was the original October 2024 project and its token. In January 2025 the team rebranded the software and DAO to elizaOS after Andreessen Horowitz asked for distance from the AI16Z name. A token migration later moved most holders from the legacy AI16Z token to a new ELIZAOS token, though a small amount of the original AI16Z token still trades separately on decentralized exchanges.

Why has the ELIZAOS token price collapsed even though the framework is still active?

The token and the open-source codebase are separate assets that have effectively decoupled. The token’s problems include a supply migration that expanded circulating tokens roughly tenfold, a federal class-action lawsuit against the project’s founders alleging the original autonomy claims were overstated, and a broader loss of speculative interest in AI-agent tokens generally. None of those factors slow down GitHub development, which is why the framework keeps shipping releases while the token keeps setting new lows.

Is elizaOS free and open source to build on?

Yes. The core framework is MIT-licensed and hosted on GitHub at no cost, and developers can self-host an agent entirely for free. The project also offers optional paid cloud services for hosting and authentication for teams that would rather not run their own infrastructure, but that layer is separate from the free, self-hosted option.

What is the status of the ai16z/ElizaOS class action lawsuit?

The case, Pikabea v. Walters, was filed in the Southern District of New York in April 2026 and, as of the most recent public hearing, had not progressed past serving the defendants. Presiding judge Jed Rakoff criticized the plaintiffs’ single attempt to serve founder Shaw Walters and ordered a written accounting of service efforts. No defendant has responded to the allegations on the merits, and none of the claims have been proven in court.

Marcus Okafor, Research Editor, HOGE Wire AI × Crypto desk.