Hey HN! We've been working on Station, an open-source runtime for what we call "operational agents" - simpler AI agents focused on ops tasks rather than full application embedding.

The key insight was that agent management shouldn't be confined to your IDE. We built Station with an SSH-first approach using wish + bubbletea for a rich terminal experience that feels natural for ops workflows.

What makes Station different:

- MCP-native: Station itself is an MCP server supporting both STDIO and SSE, working seamlessly with Claude Desktop and other MCP clients

- Tool restrictions: Assign specific tools to agents to prevent model degradation from tool overload

- GitOps-friendly: MCP configs as template bundles with proper secrets management

- Single binary: Deploy anywhere with SQLite + optional Litestream replication

- Export flexibility: Generate dotprompt files to manually tweak agent behavior

- API access for programmatic control alongside the SSH interface

We're in early beta and would love feedback from the community. The vision is to make operational agents as easy to deploy and manage as any other ops tool.

GitHub: https://github.com/cloudshipai/station

Happy to answer questions about our architecture choices, MCP implementation, or anything else!