Welcome to the world of Web3 agents! If you're interested in building and using agents locally, this document will guide you through the principles and current projects underway.
- Agents cannot execute decisions: Agents should not be given private keys or allowed to make transactions on their own. They can only construct transaction payloads for a user's approval. This is due to the limitations of current LLMs in understanding complex transactions and the risk of gaslighting.
- Local installation: Agents should run on the user's laptop, typically with 8-16 GB of RAM. This allows for faster execution and better performance.
- No private keys: Agents must not have access to private keys or be able to execute transactions independently. User's cryptographic approval is essential for any transaction.
- lachsbagel on Discord - this repo:
- Local Docker install
- (pending) HideNSeek: An algorithm for verifying and fingerprinting which model a compute provider is actually running
- Note: this repo will be made public under an MIT license following the publishing of a paper by the same name which is currently in the works
- (future) Posting Agent - an agent which optimizes content for social media posting
- IODmitri on GitHhub 3. (pending) HideNSeek: An algorithm for verifying and fingerprinting which model a compute provider is actually running
- artfuljars on Discord:
- Windows Build (EXE version of .app)
- Two installation wizards:
- Windows
- (pending) macOS
- (pending) CICD builds for Windows, Linux, and macOS
- (pending) Vulnerability scanning of dependencies and code
- GenLayer
- (pending) FeedBuzz - AI filtered logging system to surface user demand and failure modes for new functionality
- CliffordAttractor on Discord - Following Assume 16GB+ RAM:
- Developed a price fetcher agent using CoinGecko.
- A web interface which is served by the local Docker installation and integrated with Rainbow, enabling the use of MetaMask, WalletConnect, and other EVM-based wallets.
- A swap agent which can iteratively ask users to provide needed details for disambiguation.
- (Pending Integration) A general-purpose agent that can ingest arbitrary documents, such as PDFs, for basic document QA and text generation.
- (Pending) Delegating agent which can maintain user's persona/interests as well as coordinating to task agents and tools.
Pending Lumerin's work. Eventually Agent Builders will be able to permission-lessly upload Agent and Model artifacts to a decentralized registry.
- (pending) FLock is working on a decentralized agent which will regularly be updated with most recent docs surrounding the Morpheus system for the following audiences:
- Normal user perspective (broad Q/A for those new to Morpheus ecosystem)
- Developer perspective wanting to deploy on Morpheus
- Perspective of capital contributors
- If you are working on an agent which can provide value through open models and relies on processing public data, please reach out to lachsbagel on Discord (link below)
- Otherwise, you are more than welcome to publish your agent to the registry when it goes live pending Lumerin's work and any other necessary pieces which come up to better ensure security and verifiability of models in non-local execution environments.
- If you are working on security and/or verifiability of models and the runtime, please reach out to LachsBagel on the Morpheus Discord.
- Currently looking at Hyperbolic.xyz and 6079. See more ecosystem members here.
- LachsBagel is also working on a new algorithm, named HideNSeek, which uses a Transformer specific heuristic for model verification
- Atoma.Network will help with implementing the plumbing for HideNSeek
Join the Morpheus Discord
Last Updated: July 7, 2024