Golem 1.5 came out on October 9 with hands on the desktop: Claude could read the windows of another machine, write into their text boxes and press their buttons. Three more versions have come out since then. This note puts them together, because each one fixes something you noticed when using it.

1.5.1: the hands, tested on real machines

I published 1.5 before certification was finished, because a customer needed it. Certification then took it through a customer’s computer with Golem Assist, a server with the Guardian and the login screen of an ERP built in Visual FoxPro, and it found a serious bug. With two windows of the same program open, for example File Explorer and the “Run” box, which are both explorer.exe, the keys meant for one could end up in the other. Golem checked that the program was in front, not the window.

In 1.5.1 it checks the window, and it does so before every key. If somebody starts working in another window halfway through, Golem stops and says how many keys went out.

Other fixes from the same certification came with it:

  • Under the Guardian, the helper that works on the desktop can no longer write into administrator windows, including when you log in with the built-in Windows Administrator.
  • With the workstation locked (Win+L), the tools that write or press do nothing and say so. Before, a button could be pressed through UI Automation behind the lock screen.
  • What Claude writes as a secret does not come back out somewhere else. A Visual FoxPro password box does not declare itself as one to Windows, and the value could be read back in plain text; Notepad copies the text to its title, and the title ended up in the log. Now the Golem agent remembers in memory what it was sent as a secret and replaces it with *** in everything it reads and in every line of the log.
  • In look-only mode, the three tools that read the screen can be used.

1.5.2: a long call no longer cuts the connections

The Claude app cuts off a tool that takes longer than 60 seconds, and on the next call it restarts the Golem client. That lost every open connection, including those of other conversations, which share the same client. A three-minute run_powershell was enough to bring everything down.

Since 1.5.2 no call waits longer than 45 seconds. If the remote tool has not finished, Golem answers pending with an id, the call keeps running on the other machine and Claude picks it up with resume. This version only changed the client.

1.5.3: FoxAgent and the other child servers

Golem can start other MCP servers on the other machine, such as FoxAgent, and publish their tools next to its own. If one of those servers goes down, Golem starts it again and repeats the call.

Up to 1.5.2, Golem took a server that answered with an error, or that took longer than the time limit, for one that had gone down. It closed it with whatever it had open, started it again and repeated the call. Certification showed that this way a tool failing with an ordinary error ran twice. In 1.5.3 only a server that has gone down is restarted, because its process died or because the connection to it was cut. If it answers with an error, Golem returns that error as it came, and if it runs past the time limit, Golem says so without repeating the call.

Simultaneous calls also got stuck. On a server with two processors, eight calls at once to FoxAgent took 6 seconds, and with 48 six were lost. With 1.5.3, in the same test, the eight take less than a tenth of a second, and 96 calls from four clients at once all get in under 0.15 seconds.

This version changes the agent and the client. To get it on a server, its golem.exe has to be updated, or its Guardian’s.

Tested

1.5.1 and 1.5.3 went through certification on real machines before release, and 1.5.2, which only changed the client, through the usual tests. The current version, 1.5.3, is signed, and the hashes of every file are in the installation guide.