Vercel fx 0.0.16: Prompt Caching Changes and the libfx Upgrade Caveat

Vercel Labs released fx 0.0.16 on October 29. It reports 41% fewer uncached input tokens in a benchmark, not a guaranteed 41% bill reduction.

That distinction is the useful part of this release. Agent costs depend on what the harness repeatedly sends, how much the provider can reuse and which tokens are charged at each rate.

What is Vercel fx?

The fx repository describes an experimental, native coding agent written in Zig. It runs in a terminal and supports files, shell commands, web access, MCP servers and subagents.

Developers can call it from scripts with fx ask, connect editors through the Agent Client Protocol, or embed its core in JavaScript applications through libfx. The project supports macOS and Linux; its CLI does not require a separate language runtime.

That makes the audience broader than people choosing a terminal interface. Teams embedding an agent inside another product also inherit decisions about context handling, session storage and dependency packaging.

What the 41% token reduction means

The release comparison covers caching scenarios across ten model families against 0.0.15. MiniMax used 3% more uncached input tokens.

Changes preserve cacheable context across workspace changes and resumes. Skill instructions accompany their invoking turn rather than every subsequent request.

Caching efficiency is different from doing less work. An agent may need the same instructions and conversation history while avoiding repeated processing at the uncached rate. Whether that lowers the bill materially depends on the provider’s pricing, cache behavior and the rest of the task.

A sensible comparison therefore tracks completed work alongside uncached input, cached input and output tokens. A single total-token number can hide the change this release is trying to make.

Smaller packages, with a migration warning

libfx downloads shrink about 60% by selecting platform-specific binaries.

libfx requires Node.js ^20.19.0 || >=22.12.0. Version 0.0.15 cannot reopen turns with new context updates; upgrade all workers sharing a session store before new turns.

Treat the shared-session transition as part of the rollout, rather than assuming backward compatibility.

For teams evaluating fx, this release makes the agent harness itself worth measuring. Changing the model is only one way to affect cost; changing how repeated context reaches that model can matter too.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn implementation details and product updates into stories developers can understand and use.

Book a 30-minute call.