Mysti

Orchestrates multiple LLM agents (Claude, OpenAI, Gemini) in a brainstorm-and-debate loop where each agent proposes solutions to coding problems, critiques alternatives, and a synthesis agent selects the best approach. Uses agentic workflow patterns with turn-based message passing and structured reasoning to converge on optimal code solutions rather than relying on a single model's output.

Integrates agentic code generation directly into VS Code's editor as a native extension, allowing developers to invoke multi-agent workflows on selected code or cursor position without leaving the editor. Preserves editor context (open files, selection, cursor position) and streams agent responses back into the editor with syntax highlighting and diff visualization for code insertions.

Manages agent lifecycle across multiple LLM providers (OpenAI, Anthropic Claude, Google Gemini) with automatic fallback routing if a provider fails or rate-limits. Routes different agent roles (brainstormer, critic, synthesizer) to different models based on provider availability and configured preferences, with built-in retry logic and provider health checking.

Implements context management for multi-agent workflows by allowing developers to explicitly include/exclude files and code snippets in the agent context window. Uses file tree selection UI in VS Code to build a curated context set, with intelligent truncation and summarization of large files to fit within token limits while preserving semantic relevance for agent reasoning.

Captures and displays the full debate transcript between agent instances, showing each agent's proposed solution, critiques of alternatives, and the synthesis reasoning for the final selected approach. Renders debate history in a structured panel with collapsible agent turns, allowing developers to understand why agents converged on a particular solution and what trade-offs were considered.

Enables developers to describe coding problems in natural language ('vibe') rather than formal specifications, with agents interpreting intent and generating solutions that match the described vibe. Uses multi-agent interpretation to disambiguate natural language intent and synthesize code that aligns with the developer's described approach or style preference.

Assigns specialized roles to different agent instances (brainstormer, critic, synthesizer) and routes each role to the LLM model best suited for that task. Brainstormers use creative models, critics use analytical models, synthesizers use reasoning-optimized models, with configurable role-to-model mappings allowing teams to customize agent specialization based on their model preferences.

Implements iterative refinement where developers can request agents to improve generated code based on specific feedback (performance, readability, security, style). Agents use feedback to generate revised code and explain what changed and why, with multi-agent debate on refinement approaches to ensure improvements address feedback without introducing regressions.

Provides open-source TypeScript codebase (1037 GitHub stars) enabling developers to extend agent behavior, add custom agent roles, implement alternative debate strategies, and integrate with additional LLM providers. Uses modular architecture allowing community contributions for new agent types, reasoning patterns, and provider integrations without requiring core maintainer involvement.