fx v0.0.17: Vercel Labs coding agent adds subagents, Token Pass

Vercel's tiny fx coding agent gets subagents you can steer

Summaryfx v0.0.17, the open-source coding agent from Vercel Labs, adds subagents you can switch to and message, built-in Vercel Token Pass models, and tool definitions 48% smaller. It is still labeled experimental. Not tested by MODEL.

November 14, 2026 · MODEL Editors · 2 min read

Some links may be affiliate links. Disclosure

What it is. fx is a coding agent harness and command-line tool from Vercel Labs, written in Zig. Its site describes it as open source (Apache-2.0), model-agnostic, and built to run with local or cloud models, shipped as a single 6.33 MiB binary (fx.sh (opens in a new tab)). Its output is meant to feel closer to a Unix shell than an IDE in the terminal. Source is on GitHub (opens in a new tab).

What's new in v0.0.17. Released Saturday, Nov. 14 (changelog (opens in a new tab)):

Subagents work alongside you. They report back when they finish, and you can switch to any subagent to watch it, message it or answer its prompts. Each one gets its own model, effort level and permission mode, and they can nest two levels deep.
Agent profiles. Files in ~/.fx/agents or .fx/agents set an agent's tools, MCP servers, model and prompt; fx --profile <name> starts a session with one.
Several instances in one window with fx --instances <n> or /launch.
Vercel Token Pass is built in. Token Pass models come first in the model picker, and /credits shows your Token Pass credit and its reset date.
Smaller requests. Vercel says built-in tool definitions are 48.0% smaller than in v0.0.16, and in its own live tests prompt-cache reads rose from 61.3% to 82.0% of later input.

Who it's for. Developers who want a small agent for the terminal, CI or a sandbox, or who want to embed one in their own app (the libfx library also runs in the browser via WebAssembly).

The catch. The fx site labels it "experimental" and says "use at your own risk, we will be making frequent changes." The size and cache numbers are Vercel's own measurements. Token Pass is a Vercel product, so the built-in model picker leans toward Vercel's own offering, though fx still works with other providers and local models.

The MODEL take. Subagents you can actually switch into and steer are the interesting part. Worth a try on a side project; not something to pin a team workflow on at version 0.0.17.

Not tested by MODEL. Not sponsored. No affiliate links.

Keep reading

Hot take

65% of committed code is now AI-assisted. Only 39% say it's always checked.

Sonar's new State of Code survey of 1,026 developers and engineering managers says AI output is outrunning review. Our take: the bottleneck moved from writing code to checking it.

November 14, 2026 · MODEL Editors

Custom themes

Claude Code desktop gets custom themes: recolor the app with a mod

Anthropic says you can now build custom themes for Claude Code on desktop with mods: change the colors, add an animated background, or just describe the look and let Claude write it. Local desktop sessions only for now. Not tested by MODEL.

November 14, 2026 · MODEL Editors

Projects update

Claude Code Projects update: a more proactive coordinator, /btw, /fork and reminders

Anthropic's Claude Code Projects now has a coordinator that answers in the project chat and reports on its parallel threads unprompted, plus /btw, /fork and thread reminders. Projects is in public beta on Pro, Max and Ultra. Not tested by MODEL.

November 13, 2026 · MODEL Editors

Now in the API

DeepSeek puts its chat app's search, file reading and voice into the API

Web search, deep thinking search, file input, URL reading, safety checks, OCR, speech to text and text to speech are now available to developers through the DeepSeek API, plus an open-source demo chat app. Not tested by MODEL.