Loomcycle, a self-hosted AI agent runtime, releases v1.0 with hardened features and multi-language integration—offering teams an alternative to embedding agent libraries or renting cloud services.

What happened: Loomcycle 1.0 is now released as a standalone Go binary (~50 MB) that runs alongside your application and coordinates AI agent loops. It stabilized core features across v0.8–v0.23 (multi-replica HA, agent/skill definitions, MCP on both sides, memory backends, webhooks) and then hardened the runtime v0.24–v0.37 with interactive terminals, pause/resume, context compaction, and robustness for slow local models. The v1.0 milestone itself is a pure hardening release with no new primitives, distributed via Homebrew, multi-arch Docker, and the Claude Code plugin, and passed an 8-hour stability soak running 1.27M circuits and 3.8M agent runs with 100% completion and zero leaks.

Why it matters: Teams that want to run AI agents can today either embed a library inside their application, rent a managed cloud service locked to one vendor's infrastructure, or proxy through a gateway that doesn't actually execute agents. Loomcycle offers a fourth path: a lightweight self-hosted runtime you control, written in one language while your application stays in whatever language you chose, and compatible with HTTP, gRPC, MCP, TypeScript, and Python interfaces.

What to watch: The software is Apache-2.0 licensed and the team welcomes bug reports, security disclosures, feature contributions, downstream consumers, and forks (see CONTRIBUTING.md). It supports seven inference modes (Anthropic, OpenAI, DeepSeek, Gemini, Ollama local and cloud, plus a synthetic code-js provider) and includes 19 built-in tools with Claude Code parity plus vector memory on pluggable backends.