Agents In Use
What’s new todayLast update
3 use cases added today
News / FIELD GUIDE

Hermes 0.23 brings browser, bots and cron to one gateway

Hermes Agent 0.23 adds browser access, opt-in bot routing and controls for scheduled work. What the release records show, and what still needs your checks.

By Agents In Use editorial6 min read

Hermes Agent's November 5 release puts its Desktop workspace in a browser tab and gives each machine's gateway ownership of its conversations. The 0.23 release notes describe the CLI, Desktop, messaging bots and cron attaching as clients instead of running separate agents against the same chat. For someone moving between a terminal and a messaging app, the intended benefit is one queue and one history, rather than competing copies of the conversation.

That is a useful operational change to examine separately from model capability. Our reading of this release is that conversation ownership matters more than another way to open a chat window. The accompanying implementation records also describe opt-in routing between bots, persistent connections to multiple gateways and controls for stopping scheduled work. These are maker claims and development records, not an independent test by Agents In Use.

One conversation owner is useful only if you can also see who is working and stop the work.

A browser window into your own gateway

The Webapp serves the Desktop workspace from the user's gateway, including chats, Group Chats, plugins and settings. The release notes say it is off by default. Users enable it through Web Access in Settings or the configuration setting webapp.enabled. Browser access is therefore an option to turn on, not a claim that upgrading automatically exposes a workspace.

The Webapp implementation, merged on October 25, is older context for the November release. It describes the gateway hosting the browser workspace without a separate Webapp server or relay. It also distinguishes local access from access on other machines, where a sign-in gate applies. Those details matter more than the convenience of another tab.

Our recommendation is to start with local access and check the sign-in behaviour before configuring a public address. Decide which machine owns the work before opening the same task through several clients. Treat browser access as an access-control decision, not permission to widen the agent's tools or credentials.

Connected machines are still different machines

The release also keeps registered gateways connected in Desktop. The multi-gateway change, merged on November 1, says switching gateways no longer requires the old confirmation dialogue or clears the session list. Its implementation opens the target before changing the current view, so a failed target connection leaves the current session and draft untouched.

The author's verification used real Electron builds and real backends, including token-based remote access and an SSH gateway. The record explicitly excludes live testing on Windows and macOS hosts and OAuth or cloud gateways. That is a useful limit to retain: a developer's successful setup is evidence for that setup, not every supported connection.

For a laptop and home server, our advice is to give the gateways unambiguous names and check the selected host before a file-writing task. The release notes say terminal tabs run on the host where they were opened. Keeping connections alive should not be mistaken for moving files, permissions or running processes between computers.

A lead bot can narrow who answers

Group Chats gain an optional lead that chooses who should handle an unaddressed message. The lead-routing record, merged on November 3, says existing rooms keep their default behaviour, where everyone gets a turn. Selecting a lead is a per-room choice, not an automatic change to every conversation.

The record describes a visible assignment line when the lead hands work to another member. Explicit mentions bypass the lead. If the lead returns an invalid choice or fails, the lead itself takes the message rather than broadcasting it. Scheduled posts and replies inside a thread keep their existing resolution behaviour.

The testing distinction is important here. The author exercised the real Electron app and backend with three bots, but used mocked inference for the routing decision. That supports the described routing mechanics. It does not demonstrate that a model will consistently choose the right specialist for an ambiguous real request.

Scheduled work gets a stop control

The cron control change, merged on October 30, adds a Desktop Stop button and hermes cron cancel. It also lets each job choose how to handle runs missed while the machine was off. The documented options distinguish running once when back from skipping missed work, with the configured default available too.

The implementation record separates cancelling a live run from pausing its schedule. It shows a running job that is already paused still offering Stop and Resume. For operators, our recommendation is to make that distinction explicit in runbooks: do you want to halt the current task, prevent future runs, or both?

Choose the missed-run policy task by task. We would usually skip a time-sensitive briefing after its usefulness has passed, while considering a catch-up run for a reconciliation task whose inputs remain relevant. These are editorial examples, not policies the release chooses for you. Check the run history rather than assuming an offline period produced the result you expected.

What the release does not establish

Two historical accounts help explain the work these controls might support. In the October 13 Grok Bot and Hermes bridge account, the poster describes document requests between a cloud assistant and a local agent. The case records experimental limits, including missing delivery acknowledgements. It is not a test of this release or an official integration.

The August 7 Notion meeting workflow describes a user reviewing Whisper transcripts before Hermes prepares summaries and linked actions. NDA restrictions prevent public inspection of the actual meeting records. Neither account establishes that a shared gateway makes the underlying work accurate. They illustrate why retaining a human review step remains sensible.

There is also a newer maintenance detail to check before adopting 0.23. The 0.23.1 notes report a fix for clients rejecting an oversized live plugin catalogue and falling back to their bundled catalogue. The notes say older clients retain that fallback until updated. Our recommendation is to verify the installed version and plugin availability as part of an upgrade check, rather than assuming the 0.23.0 announcement describes every subsequent fix.

Sources