Skip to content
GitHub

HotLoop Flow 2.14.0 release notes

HotLoop Flow 2.14.0 lets an agent reach the rest of Flow over MCP, and drains a backlog after an outage in batches instead of one message at a time. It adds 22 MCP tools, a page that lists all 29, and a faster way for a durable wire to catch up on an InfluxDB that has been away. It is amd64 first, with arm64 following in 2.14.1.

  • MCP tools for the rest of the admin API. Flow’s own tools covered the flows and nothing around them: list, read, deploy, roll back and press an inject node. There are 22 more, so an agent can read the settings, the palette and the plugins, read the audit log, read a deployment and what changed between two, diff and type-check a draft against what is running, export the flows, read, replace and add flow tests and run them, read the message traces and draft a test from one, replay a trace, read the runtime’s numbers, and turn off a user’s second factor (reset_mfa, which needs auth.admin). Each tool is one admin route run through its own handler, with the permission that route asks for, so a refusal comes back in the route’s own words and every call is in the audit log under the caller’s name. Sign-in, a person’s own second factor and the live editor socket are not tools, and the MCP page says why. That page lists all 29 tools with their routes and permissions.
  • A durable wire drains in batches. After an outage, influxdb out writes a backlog up to 256 messages to a request instead of one. With a historian that takes 8 ms to answer a write, a plant sending 100 readings a second and 2,000 readings queued, the same test drained in 2 minutes 38 seconds before and in 2.7 seconds now, every reading once and in order.
  • A durable wire to a node that takes batches claims up to 256 messages at once. If a batch is not taken, the messages before the first one that failed are marked done, that one is retried on its own with the retry and give-up the wire has always had, and the rest are sent again, so nothing is lost or reordered. Every other target still gets one message at a time. Delivery is still at-least-once, and a message that was in flight when the process was killed can come again as before. For a batched target that is up to a batch, up to 256 messages, rather than one, and the _queuedAt stamp makes a repeat to InfluxDB an overwrite as it was.
  • The LICENSE file’s Licensed Work line names only HotLoop Flow, because the other products move to their own license. The terms are not changed.
  • Nothing in the editor changed.
  • Nothing else behaves differently from 2.13.1.

None.

A plain helm upgrade to the 2.14.0 image and chart, or the new tag in your Quadlet unit. Nothing happens on first start and nothing needs migrating. An agent that connects to /mcp sees the new tools and only the ones its token’s permissions allow. This release is amd64 only: arm64 boxes stay on 2.13.1 until 2.14.1, which is the same code with the arm64 image added.

Image: ghcr.io/hotloop-io/hotloop-flow:2.14.0, linux/amd64 first, with arm64 following in 2.14.1. Chart: https://hotloop.io/hotloop-flow/. The README at this tag has the rest, with pictures: https://github.com/HotLoop-io/hotloop-flow/blob/v2.14.0/README.md