Moonshot AI’s Kimi Code development release bundles substantial changes to the coding agent, including more deliberate control over plans and tool approvals, plus new remote-control features. Its release notes also describe a separate development release channel, so this is a preview of where the tool is heading rather than a claim that every change is a stable release.
Moonshot AI/Kimi Watch analysis
What happened
The GitHub release for moonshot-ai/[email protected] carries a long changelog spanning Kimi Code versions 2.1.1 through 2.3.3 and subsequent development work. Among the newer additions, users can launch the desktop app from the CLI, manage remote-control devices, and transfer files in chunks of up to 100 MiB.
The notes also say Kimi Code now leaves Plan mode for the user to exit and reviews every plan, supports symlinked directories in its file tree, and has an independent development release channel. Changes to approvals and questions make them available as interaction frames in remote control. This is a broad set of changes, not one neat feature tucked into a tidy little release.
Our top picks
- Plans stay with the user
The agent leaves Plan mode for the user and reviews every plan, adding a clearer checkpoint before work proceeds. - Remote control gets practical tools
Device management and chunked transfers up to 100 MiB extend what users can do away from the local session. - Approvals become visible interactions
Remote-control sessions can receive approvals and questions as interaction frames. - A separate development channel
The new channel gives users a way to try development builds apart from the main release stream. - Symlinks work in the file tree
Support for symlinked directories removes a small but familiar obstacle in real-world projects.
Why it matters
Coding agents are increasingly expected to work across sessions, tools and remote machines, not just produce a plausible block of code. These changes address that practical terrain, while the plan-review and approval features put more visible checkpoints around the agent’s work.
The changelog is an aggregate across versions, so it should not be read as a list of features all newly introduced on the same day. The development tag matters too: keen users may welcome the preview, while teams relying on stable behaviour may prefer to wait.
Our read
There is useful substance here, especially in remote control and the added user checkpoints. But this is a development release with a wide, accumulated changelog, not a single polished launch moment. Try it if you want to test the direction of travel; do not treat a preview tag as a quiet invitation to upgrade production tooling.
What to watch
- Which changes make it into a stable release, and when.
- Whether the new remote-control features work reliably across devices and larger transfers.
- How the plan-review and approval checkpoints behave in everyday coding workflows.
Discussion spark: Would you trust a coding agent more with visible plan reviews and approval steps, or do those checkpoints just move the friction somewhere else?
Sources and evidence
- moonshot-ai/[email protected] (28 October 2026, 05:44 UTC)
- Releases · unslothai/unsloth · GitHub (28 October 2026, 05:44 UTC)
not affiliated with or endorsed by Moonshot AI and Kimi