Eliza Framework: Crypto’s Top AI Agent Stack Hits a Wall
Eliza is the open-source stack that turned AI agents into a crypto mania and a $2.6 billion token. A 99% crash, a security flaw, and a federal lawsuit now test what it built.
For a few months in late 2024 and early 2025, the fastest way to launch a crypto project was to point an Eliza agent at a social media account and let it run. Eliza, the open-source framework from Shaw Walters and Eliza Labs, became the default toolkit for building autonomous AI agents that post, trade, and manage onchain wallets. It also became the engine behind one of the strangest valuations of the cycle. The token tied to the project peaked above a $2.6 billion market value, then surrendered almost all of it. As of June 27, 2026, the same network trades as a micro-cap, a security paper has shown how its agents can be tricked into emptying wallets, and a federal class action accuses its founders of selling a fantasy. Here is what the Eliza framework is, what it built, and why it matters now.
What the Eliza framework actually is
Eliza is an open-source, TypeScript-first framework for building and running AI agents, released under the permissive MIT license. An agent is defined by a character file, a block of structured data that sets its name, personality, sample dialogue, and goals. The runtime connects that character to large language models from providers such as OpenAI, Anthropic, or self-hosted open models, then links it to the outside world through clients for Discord, Telegram, and X. The result is a persistent bot that holds a consistent voice, remembers past conversations, and can take actions instead of only chatting.
The code spread quickly. The main elizaOS/eliza repository on GitHub carries more than 18,000 stars and 5,000 forks. By December 2024 it had become one of the most active AI projects on the platform, and supporters claimed it briefly outpaced Google’s Gemini repository in contributor activity, a milestone Delphi Digital tracked during the run-up. For a crypto-native project, that level of developer attention was unusual.
From an a16z joke to a $2.6 billion market
The project started as a wink at Silicon Valley. In October 2024, Walters launched a decentralized autonomous organization on the daos.fun platform with a raise of about $75,000. He named it ai16z, a play on Andreessen Horowitz (a16z), and gave its flagship trading agent the name Marc AIndreessen, a nod to the firm’s co-founder Marc Andreessen. The pitch was a venture fund run by AI agents that would invest a community treasury onchain, and traders treated it as the purest bet on the AI agent theme.
Speculation did the rest. The ai16z token climbed to an all-time high near $2.47 on January 2, 2025, according to CoinGecko data, lifting its market value past $2.6 billion. The token became shorthand for the entire sector, and dozens of copycat projects forked Eliza to ship their own agents and their own tokens within days of each other.
The rebrand that did not stop the slide
The name was always borrowed, and the lender wanted it back. In January 2025, Andreessen Horowitz asked the project to drop the ai16z branding. Chris Dixon, who runs the firm’s crypto arm, said on the Unchained podcast that a16z had no connection to the project and had requested the change. Walters agreed, and the project rebranded to ElizaOS, a move reported by The Block and detailed by Unchained. Walters initially said no new token would be created and that holdings and governance would not change.
The rebrand did not arrest the decline. As the wider AI agent narrative cooled through 2025, the token drifted lower, and a later migration to a fresh ELIZAOS contract reworked the supply structure. That decision, and the promises that surrounded it, would become central to the legal fight described below.
ELIZAOS by the numbers
The gap between the peak and the present is stark. The figures below come from CoinGecko and reflect the network’s standing in late June 2026.
| Metric | Value (late June 2026) |
|---|---|
| Price (USD) | About $0.00061 |
| Market capitalization | About $4.6 million |
| Fully diluted valuation | About $5.3 million |
| Circulating supply | About 7.48 billion ELIZAOS |
| 24-hour trading volume | Roughly $0.8 million |
| ai16z all-time high | $2.47 on January 2, 2025 |
| Drawdown from the peak | Around 99.9% |
In plain terms, a network that once carried a multi-billion-dollar valuation now trades for the price of a small startup. The framework’s GitHub momentum and its token price have moved in opposite directions, and that split sits at the center of the project’s identity question.
How an Eliza agent is put together
Under the surface, Eliza is modular, and that design is part of why it spread so fast. Developers assemble an agent from a small set of building blocks rather than writing one from scratch.
- Character files define identity: name, bio, lore, sample posts, and tone.
- Providers feed real-time context into the model, such as wallet balances, the time, or market data.
- Actions are the things an agent can do, from sending a token to posting a reply.
- Evaluators review conversations and update what the agent remembers.
- Plugins bundle actions and providers for a specific service, such as a blockchain or an exchange.
- Clients connect the agent to Discord, Telegram, X, or a custom application.
Memory ties it together. Eliza stores conversation history and facts in a vector database, then pulls relevant pieces back into the prompt, a retrieval setup that lets an agent reference something a user said days earlier. That persistent memory is the framework’s strongest feature. It is also, as researchers later showed, its softest target.
auto.fun and the v2 reset
Eliza Labs tried to convert open-source attention into a product business. On April 17, 2025, it launched auto.fun, a no-code launchpad that lets non-developers spin up an AI agent and an accompanying token from a browser. The team marketed what it called a fairer than fair token model, meant to route more value to creators and the protocol rather than to early snipers.
Alongside it, Walters introduced ElizaOS v2 at the CATSTANBUL event in 2025. The rewrite promised a cleaner architecture, better planning so agents can chain several steps toward a goal, and a unified wallet system to fix the fragmented key handling of the first version. The ambition was to move Eliza from a hobbyist framework toward something closer to an operating system for agents, the phrase the project now uses to describe itself.
The security flaw researchers found
In March 2025, researchers from Princeton University and the Sentient Foundation published a study of agents built with Eliza, focused on what they called context manipulation attacks. Their finding was blunt: an agent that controls a wallet can be talked into draining it.
The arXiv paper describes a technique called memory injection. Because Eliza stores history externally and does not verify the integrity of those entries, an attacker who can message the agent on one platform, say Discord, can plant a fabricated instruction in its memory. Later, when a user on X asks the agent to send funds, the poisoned memory can redirect the transfer to the attacker’s address. Since all plugins share the same memory, a single bad entry can spread across every surface the agent touches. Crypto news outlet Decrypt summarized the problem as an AI that can be gaslit into losing millions. For a framework whose main job is to hand agents control of money, that is not a corner case.
A lawsuit and a new SEC rulebook
On April 20, 2026, a proposed class action was filed in the U.S. District Court for the Southern District of New York. The lead plaintiff, a Spanish investor named Gorka Pikabea, alleges that the founders marketed an autonomous AI agent that humans actually operated, pointing to October 2024 reporting from the outlet Protos. The complaint, covered by Cryptopolitan, also challenges the token migration that allegedly expanded supply roughly tenfold, from about 1.1 billion to 11 billion units, with a large share routed to insiders rather than existing holders. It identifies thousands of wallet addresses that lost money over the class period. These are allegations; the defendants have not been found liable, and the case is at an early stage.
The legal pressure lands as the rules for tokens like this one are being rewritten. On March 17, 2026, the SEC and the CFTC issued a joint interpretation sorting crypto assets into categories such as digital commodities, digital tools, and digital securities. Under SEC Chairman Paul Atkins, the agency’s Project Crypto has pushed for clearer lines on what counts as a security, as law firm Cooley noted. An AI agent token sold on the promise of returns from a managed treasury sits in an awkward spot under any of these tests, and the new framework gives U.S. plaintiffs and regulators sharper tools to ask where it belongs.
What to watch next
Eliza now lives a double life. As code, it remains one of the most adopted AI agent frameworks in crypto, with real developers shipping real bots on top of it. As a token, ELIZAOS is a cautionary tale about how fast a narrative can inflate and deflate. The two stories do not have to end the same way.
Three things will decide the next chapter. The first is whether auto.fun and v2 produce agents people use for reasons other than speculation. The second is whether the team and the wider field close the memory and prompt-injection holes that make wallet-controlling agents dangerous. The third is how the New York case and the SEC’s new taxonomy treat tokens marketed on the promise of autonomous, money-managing software. For builders watching the AI-crypto cluster, Eliza is still the reference implementation. For investors, it is a reminder that a popular repository and a sound token are not the same thing.
By the HOGE Wire editorial desk, covering the AI-crypto cluster.