Msty Claw 0.4.0 is out.

Msty Claw 0.4.0 expands local-first workflows with direct bot usage in local conversations

This is the second major Claw release in about a week, which is fun and slightly unhealthy.

The release is not just more buttons. Most of the work is about keeping agent sessions usable for longer:

  • Local bot chat: bots can now run directly in local conversations instead of starting from a remote channel.
  • Manual compaction: there is a UI control and a /compact command for cutting context when you decide it is time.
  • Caveman mode: shorter agent responses, fewer tokens, less noise.
  • Better Windows support: command execution, tool calls, PowerShell integration, and success/failure reporting all got better.

The Windows work was probably the least glamorous and the hardest.

It sounds simple until you try to make an agent harness behave well across Windows shells, approvals, long-running tool sessions, PowerShell exit behavior, and native network commands. Every operating system has its own little opinions. Windows has a committee.

This release does not mean that work is finished. It means the loop is less painful, and we can keep pushing it forward.

The token-saving work is the part I personally feel every day.

Caveman started as a funny idea:

why use many token when few do trick

I use it more than I expected. Agent sessions get wordy fast, and those words are not free. They cost tokens, attention, and sometimes the useful parts of the context window.

Installing and configuring Caveman by hand is a small chore, so we made it one click inside Claw. It now sits next to output distillation and compaction under the broader Context Saver work.

My preferred agent output is not always “beautiful.” A lot of the time I want:

what changed
what failed
what to do next

If the agent can say that in half the tokens, I will take the half.

Full release notes: Msty Claw changelog.