Skip to content
GitHub

HotLoop Cascade 1.1.0 release notes

1.1.0 gives Cascade an MCP server, and the screens the editor was missing, and moves sign-in into the editor. The engine, the workflow file format and the REST API are what they were in 1.0.2.

It’s at /mcp, on the same address as the API, and it ships on and it writes. Every MCP in the HotLoop family reads and writes, and Cascade was the last one with no MCP at all.

An agent signs in with an API key and gets what that key’s holder could do, cut down to the key’s scopes. There are 79 tools, and each one is a route of the REST API: a call goes through the route’s own handler, so the permission check, the validation, the write and the audit row are the route’s. Nothing was copied into the MCP server, so nothing can drift, and nothing was left out of it, so nothing is skipped.

  • A tool that changes something wants a reason, and the audit log keeps it next to the change, with the key and the name the client gave itself.
  • A credential’s secret never comes back, to anyone. A credential made over MCP has its secrets sealed as they arrive.
  • HOTLOOP_CASCADE_MCP and HOTLOOP_CASCADE_MCP_WRITES are the deployment’s switches, both on by default. No tool and no API route can change them, so an agent that can edit the instance’s settings still can’t turn writes back on.
  • The routes that turn a key away themselves (making API keys, SCIM tokens, secret providers and log destinations) are not tools, because a key that could make keys is a leaked key that keeps itself alive. the MCP page lists them with every tool and the permission it asks for.

Credentials, Projects, your Account, and for owners and admins Settings, Users and the audit log. Each follows the same permissions as the API: a person is only offered what their role allows, and the server still refuses the rest.

  • Credentials are made, opened, changed, shared with other projects and deleted. A secret goes in once and the screen only says it is set.
  • Projects has members and their roles, and the service account keys a standing integration signs in with, shown once.
  • Account makes and revokes your API keys, which is what the API and MCP sign in with, and changes your password.
  • Users makes people and turns them off or on. Audit log is read, filtered, and checked for tampering. Settings has the second factor requirement and the error workflow.

Sign-in is the editor’s own page now, at /login, in the editor’s colors, light and dark, on a phone as well as a desk. It does everything the old page did: first run setup, passwords, directory accounts, single sign-on, passkeys, the second factor and recovery codes.

This release builds amd64 only. Multi-arch ships in its own run, so the release after this one is a single small version bump that builds both. Until then stable, latest and the short tags list this amd64 image beside the arm64 image they already pointed at, so a box on arm still pulls. The version tag, 1.1.0, is amd64 alone.

No migrations. helm upgrade to 1.1.0, or let :stable roll the Quadlet units. Putting 1.0.2 back is safe, because nothing in the database changed. What a rollback can’t undo is what an agent already did: its changes stay made, and the audit rows it wrote stay in the log with the actor type mcp. A backup before any update is still the right habit.