Hypha
An open-source TypeScript AI agent framework with local and remote adapters, policy controls, checkpoints and replay. Licensed under Apache 2.0.
Hypha is a local-first TypeScript framework for developers building AI agents that need controlled tool use, recoverable workflows and an audit trail. It separates the agent's reasoning and planning from the execution layer that enforces policies, approvals and workflow boundaries. It's open source under Apache 2.0.
The Agent Core handles ReAct reasoning, tool selection, memory and model context. The Production Harness controls execution through finite-state workflows, with checkpoints, cancellation, recovery and human review. Tools, MCP calls, memory operations and external writes pass through the same policy and tracing boundaries.
Product behavior lives in Domain Packs, which define tasks, prompts, capabilities and output contracts. Developers can reuse the runtime across different agent applications while changing their workflows and policies. Versioned contracts and runtime validation help keep application code independent of storage and execution providers.
Local adapters use SQLite, files and local processes; Docker and remote HTTP execution adapters are also available. Memory choices include Hypha Native Memory, local Native Lite and self-hosted Mem0 OSS, alongside cloud adapters for Mem0 Platform and Vertex AI Memory Bank. Where work and memory run depends on the selected providers.
Hypha records authoritative events for replay, audit and evaluation. Its caches reuse validated reasoning, tool results and memory context, but can't authorize actions or replace execution evidence. An Express API server and an example CLI use the same runtime.
