Moonshot AI has released Kimi Code 0.43.0, giving compatible models a way to load MCP tools only when needed. The release also strengthens multi-agent Tower sessions, compaction recovery and Remote Control, making this a practical infrastructure update rather than a fresh coat of interface paint.
Moonshot AI/Kimi Watch analysis
What happened
Kimi Code can now mark an MCP server as deferred, keeping its tools out of the model’s initial tool list and loading them later through selecttools. This requires an official Kimi Code model and service that declare support for dynamic tool loading, so it is a capability-aware path rather than a universal new default.
The official Kimi Code 0.43.0 release notes also describe reliability work across Tower, long sessions and Remote Control. WittyWires has not independently tested the release.
Our top picks
- MCP tools on demand
Compatible models can defer whole server toolsets, reducing the top-level list and retrieving tools only when required. - Compaction gets a ceiling
Operators can set the maximum total attempts for a failing compaction request, with five attempts as the default. - Background waits become interruptible
New steering messages can interrupt waits for background tasks, giving users a quicker route out of an agent cul-de-sac. - Tower’s foundations are tightened
Reliability fixes cover multi-agent messaging, worktrees and the review-to-merge gate. - Remote Control sheds friction
The release fixes uploads above roughly 3.5MB, reuses unchanged assets and compresses tunnel responses with gzip. - Long sessions get smoother
Web fixes target streaming stutter, unresponsive rendered content and sessions snagged by brackets or backslashes.
Why it matters
Large tool inventories consume context and make selection harder. Deferred MCP servers offer a cleaner model: expose the catalogue, then fetch the relevant tools when the job actually calls for them. The qualification matters, since the model and service must support the capability, but the direction is useful for agent setups with several sprawling integrations.
The less glamorous fixes matter too. Multi-agent coding stops feeling clever very quickly when compaction loops, background waits or worktree hand-offs misbehave. Version 0.43 concentrates on those seams, where long-running agent systems usually acquire their collection of tiny haunted cupboards.
Our read
This is a worthwhile upgrade for teams already using Kimi Code’s MCP, Tower or Remote Control features. Dynamic tool loading is the headline change, but the strongest case for updating is cumulative: more controllable compaction, interruptible waits and fewer ways for the web interface to tie its own shoelaces together.
Enable deferred MCP servers selectively and verify that the chosen model service advertises support. Existing inline exposure remains the sensible fallback where it does not.
What to watch
- Whether deferred loading measurably reduces context use and tool-selection errors.
- Which Kimi Code model services expose dynamic tool declarations in practice.
- Independent reports on Tower’s worktree and merge reliability.
- Whether on-demand MCP loading becomes the default for large agent configurations.
Discussion spark: Which MCP servers would you defer first, and what evidence would convince you that on-demand loading improves agent reliability?
Sources and evidence
- moonshot-ai/[email protected] (14 September 2026, 12:14 UTC)
not affiliated with or endorsed by Moonshot AI and Kimi