elizaOS

elizaOS is a TypeScript framework for building autonomous AI agents with persistent personality, multi-platform presence, and onchain capabilities. Agents are defined through character files (JSON personality configs), extended through a plugin system (actions, providers, evaluators), and deployed to Discord, Telegram, Twitter, Farcaster, and custom interfaces. The framework includes a RAG knowledge system, memory management with vector similarity search, trust scoring for transaction safety, and native Solana wallet integration via plugin-solana.

Source: https://github.com/elizaOS/eliza

What You Probably Got Wrong

LLMs have stale training data. These are the most common mistakes.

• "Install with npm/yarn" --> The recommended package manager is Bun, not npm. The CLI is @elizaos/cli and you install it globally with bun i -g @elizaos/cli. npm works but Bun is required for the monorepo and strongly recommended for all elizaOS projects.
• "Use npx create-eliza-app" --> This was the v1 bootstrapper. In v2, use elizaos create after installing the CLI globally. The old create-eliza-app command is deprecated.
• "Plugins live in the main repo" --> As of v2, plugins are split into their own repositories under the elizaos-plugins GitHub org. Install them from npm (@elizaos/plugin-solana) — do not copy plugin code from the monorepo.
• "Characters are in the main repo" --> Characters have been moved to https://github.com/elizaOS/characters. The main repo ships minimal example characters only.
• "Just write a prompt and you have an agent" --> elizaOS agents are not prompt wrappers. A character file defines bio, lore, message examples, style guidelines, topics, and adjectives. The runtime uses these to construct context windows dynamically. Skipping message examples produces generic, personality-less responses.
• "Actions are just functions" --> Actions must implement validate (should this action run?), handler (execute), and examples (few-shot for the LLM to learn when to invoke). Missing examples means the model will never trigger your action.
• "Memory is automatic" --> The memory system stores conversation history and knowledge embeddings, but you must configure the database adapter (SQLite for dev, PostgreSQL for production). Switching embedding models without clearing the database causes vector dimension mismatches.
• "Trust scoring prevents all bad trades" --> Trust scoring evaluates token safety using recommender credibility and token performance data, but it is a heuristic — not a guarantee. Always set transaction limits and review trust thresholds.
• "Use pnpm/yarn for the monorepo" --> The monorepo uses Bun workspaces. Running pnpm install or yarn install will fail or produce broken lock files.