3.7k
Connect

Blog / News

Radio Free Flox for May 23rd, 2026

Flox Labs23 May 2026

This week we’re debuting a new feature on the Flox Blog: Radio Free Flox. Each week, we’ll descant on the new packages, environments, AI tools, etc. that land in the Flox Catalog or go live on FloxHub. We'll also dish on what’s going on internally here at Flox.

Radio Free Flox is proudly uncensored: We intentionally publish after Flox management has signed off for the weekend. (So ... after 9 PM on Saturdays.)

This week’s episode covers new terminal coding agents; Claude Code ecosystem tools; spec-driven dev kits; MCP and browser automation integrations; servers for running frontier checkpoint models, and other goodies. We also spill tea re: the beaucoup dolares that some Flox employees spend on API tokens each month, and we explore Flox master builder Rok Garbas’ passion for procuring swine. Note: “swine” is not an agentic framework. (Maybe it should be?) It’s ... exactly what it sounds like.

This. Is. Radio Free Flox.

CLI coding harness kit

This week the Flox Catalog added packages for Amp, DeepSeek TUI, GitHub Copilot CLI, Gemini CLI, Grok CLI, Mistral Vibe, goose-cli, Qwen Code, and Forge. Some of these (Amp, GitHub Copilot CLI, Gemini CLI) are already indexed by the Catalog. Others (Cursor Agent, Mistral Vibe) are new. In every case, Flox now tracks upstream releases and publishes updated packages within hours.

Most of these tools are so well-known as not to need an introduction. For the others: this guide shows how to run the flox/deepseek-tui package. flox/mistral-vibe is another new package. It’s a credible Claude Code / Codex CLI / Gemini CLI competitor, but with a hackable (in a good way) Python implementation, plus an emphasis on configurability vis-à-vis agents, MCP, ACP, and skills. Also worth checking out.

Claude Code-adjacent kit

The Flox Catalog added ccstatusline, ccusage, claude-code-router, claudebox, and Anthropic’s experimental sandbox-runtime.

New Claude Code plugins and skill packs include:

Here’s one of our contributor's (Steve Swoyer’s) updated Claude Code manifest with flox/ccstatusline, flox/ccusage, flox/sandbox-runtime, and (why not) flox/claude-code-plugin-caveman added to it:

[install]
## core cc package
claude-code.pkg-path = "flox/claude-code"
claude-code.pkg-group = "claude-code"
 
## anthropic srt: experimental sandbox env for cc
sandbox-runtime.pkg-path = "flox/sandbox-runtime"
sandbox-runtime.pkg-group = "sandbox-runtime"
 
## caveman plugin: make cc talk like groot. save tokens? 
claude-code-plugin-caveman.pkg-path = "flox/claude-code-plugin-caveman"
claude-code-plugin-caveman.pkg-group = "plugins"
 
## ccstatusline: display cc model info, git branch, token usage, other metrics
ccstatusline.pkg-path = "flox/ccstatusline"
ccstatusline.pkg-group = "adjacent"
## query cc token usage + cost
ccusage.pkg-path = "flox/ccusage"
ccusage.pkg-group = "adjacent"
 
## useful deps
openssl.pkg-path = "openssl"
openssl.pkg-group = "openssl"
openssl.outputs = "all"
 
## flox mcp server
flox-mcp-server.pkg-path = "flox/flox-mcp-server"
flox-mcp-server.systems = ["aarch64-darwin", "x86_64-linux"]
flox-mcp-server.pkg-group = "mcp"
flox-mcp-server.outputs = "all"

Steve can flox activate -r flox-labs/claude-code this environment and get these tools anytime/anywhere. If he needs a latest/greatest Claude Code-adjacent tool that just got picked up by the Flox Catalog, he can add it declaratively (by using flox edit, with built-in linting) or install it imperatively:

flox install flox/ccstatusline flox/ccusage flox/sandbox-runtime flox/claude-code-plugin-caveman
✔ 'ccstatusline', 'ccusage', 'sandbox-runtime', 'claude-code-plugin-caveman' installed to environment 'claude-code'

Note: ccusage is machine-specific: If you work on several different machines, you’ll get different results for each machine. It also seems to calculate based on projected API usage. The code block below shows how much Steve has abused the Claude Code API on just his main dev box:

┌────────────┬───────────────┬─────────────┬────────────┬─────────────┬─────────────┐         
│ Date       │ Agent         │ Models      │      Input │      Output │  Cost (USD) │    
│ Total      │               │             │  1,390,747 │  11,368,012 │    $6331.69 │
└────────────┴───────────────┴─────────────┴────────────┴─────────────┴─────────────┘

The rest of Flox Labs—namely, Rok, Morgan Helton, and Brendan Magee—refused under threat of termination to make public their actual monthly API costs.

So maybe a $200-a-month Claude Max plan is close to a bargain? But what happens when Anthropic, OpenAI, Google, et al start charging for what this stuff actually costs them?

Spec-driven / workflow-oriented agentic development kit

The Flox Catalog added plenty of packages for spec- and workflow-driven development. Stuff like flox/spec-kit, flox/openspec, and flox/workmux aren’t new to the Catalog, but they’ll now update automatically with new releases.

flox/worktrunk is new: similar to workmux, it’s a CLI tool that aims to make Git worktrees easier to use, especially with multiple agents in parallel. Also like workmux, worktrunk is written in Rust.

Flox customers frequently ask about using Git worktrees. Both workmux and worktrunk are two credible options for doing this.

Agentic cat-herding kit

CodeRabbit CLI, isn’t new—the Flox Catalog has shipped flox/coderabbit-cli for months now—but as of this week (thanks to the Rhadamanthysian Rok Garbas), the version you get from the Flox Catalog will basically always be up-to-date.

flox/mcporter is, however, new to the Catalog. It’s a TypeScript runtime and CLI that discovers, wraps, and exposes MCP server tools as callable local CLI commands or typed APIs. Yep: it lets you call MCP servers from bash, Makefiles, scripts, or agents. Useful.

flox/agent-browser, from the folks at Vercel, gives agents a browser-control layer with screenshots, DOM and accessibility context, forms, storage, isolated sessions, and other conveniences. It aims to make browser automation more reliable for coding agents as they build and test login flows, scrape pages, capture screenshots, or script web-app interactions.

AI model-serving kit

Also “newish” in the Flox Catalog are model-serving platforms like flox/lmstudio and flox/ollama. We’ve been shipping flox/lmstudio for a couple of months, but going forward the Flox Catalog should always pick up the latest upstream version within hours of release. You can read about how to run and use it in this walk-through article.

Ollama has been available in the Catalog almost since its inception; at some point, however, the upstream nixpkgs recipe Flox was using to build the Nvidia CUDA-accelerated flox-cuda/ollama package changed, breaking compatibility on non-NixOS systems. Flox took this over with a custom flox/ollama-cuda package a few months ago, but keeping this package current on x86 and ARM was being done manually by one person. (Steve) Now both the ollama (for non-CUDA platforms) and flox/ollama-cuda packages update almost as soon as new releases ship. Obligatory disclaimer: Flox first needs to build these packages. flox/ollama-cuda, in particular, is a demanding build—it takes ~40 minutes on a 32-core/64-thread Threadripper and several hours on an emulated ARM VM—so expect a lag time of a few hours.

oMLX is a promising newish model server for macOS. It uses Apple’s open-source MLX machine-learning array framework for Apple Silicon to run language models faster than just using the Apple Metal/MPS API alone. What’s fascinating about oMLX is that it splits huge models between Apple’s unified memory (the fast memory that’s shared between the Apple M-series CPU and GPU) and high-speed PCIe 4 or PCIe 5 NVMe storage. oMLX keeps the active model's state in RAM but spills older cache data to NVMe, so oversized models are usable on Mac hardware. The flox/omlx package bundles both oMLX and the HuggingFace CLI, which oMLX uses to download checkpoint models.

Other neat kit

flox/fnox is a command-line tool for managing passwords, API keys, and other secrets, including workflows that securely sync secrets via Git or third-party password managers.

Nvidia SkillSpector scans AI-agent skills for risky or malicious patterns before you install them. Think of it as the equivalent of antivirus + static analysis for AI workflow plugins. Our flox/skillspector package delivered same-day support for this one.

flox/herdr is a is a terminal workspace manager for coordinating or multiplexing multiple AI coding agents at the same time, with support for split-screen sessions so you can track what each agent is working on. flox/serena is an open-source MCP toolkit that gives AI coding agents an IDE-style view of your codebase, so they can navigate, edit, refactor, and analyze code at the symbol level.

Finally, flox/temporal and flox/temporal-cli add workflow orchestration; they’re useful when agent systems or developer automations need long-running, retryable, and/or observable workflows.

New Flox Runtime Environments

One set of new Flox Environments centers on Claude Code plugins and agent extensions. Others address developer workflows, local AI model serving, and secrets management.

Flox environments for Claude Code skills

The flox/agentmemory environment packages persistent memory for Claude Code workflows. Test drive this environment without pulling it by running: flox activate -r flox/agentmemory.

The flox/gstack environment packages Garry Tan’s Claude Code stack, which frames agent work via role-oriented helpers such as product, design, engineering-management, release, docs, and QA workflows. Test drive this environment by running flox activate -r flox/gstack

The flox/caveman environment packages the aforementioned Caveman plugin for Claude Code. It … has to be used (and seen) to really appreciate it. Test drive this environment by running flox activate -r flox/caveman.

The [flox/remotion] environment packages skills for Claude Code and Remotion. If you use Remotion for video editing and if you’re curious, do flox activate -r flox/remotion to take it for a test drive.

Flox environments for dev workflows

Another set of new Flox environments focuses on developer workflows.

The flox/iloom environment packages iloom-cli, which is aimed at GitHub-issue-driven dev workflows where work starts from a GitHub issue or engineering ticket and the CLI keeps a structured record of plans, assumptions, and decisions. Do flox activate -r flox/iloom to take it for a no-money-down test drive.

The flox/worktrunk environment packages worktrunk with **git**, giving developers a ready workspace for worktree-based agent or parallel-branch workflows.

The flox/temporal environment packages both the Temporal server and CLI, which is useful for local development against durable workflows.

Flox environments for AI model serving

This last category is something of a grab bag. For example, flox/basic-memory is a Flox environment that packages Basic Memory, a local-first knowledge system that agents can use as a persistent memory backend. It isn't a model server, per se, but it's incredibly useful when running checkpoint models on a local model server. Sound intriguing? You can test-drive this environment by doing flox activate -r flox/basic-memory.

As for model serving:

The flox/lmstudio environment packages LM Studio for local model serving. You can read about LM Studio and model serving on Linux and macOS in this walk-through. You can take it for a test drive by doing flox activate -s -r flox/lmstudio.

The flox/omlx environment packages oMLX for local, optimized model serving on macOS. We talked about oMLX above. You can try it out by running flox activate -s -r flox/olmx.

There’s also a flox/serena environment that packages Serena as a semantic coding agent/MCP environment, including a built-in Flox-defined service for its MCP server. Try it out by running flox activate -s -r flox/serena.

Nous Research’s Hermes Agent isn’t new to the Flox Catalog, but it is new to floxenvs, the proving ground for Flox Labs. The new flox/hermes environment packages a turn-key version of Hermes. Try it out by running flox activate -r flox/hermes.

The Borgesian outlier in this heterotopic taxonomy is the flox/fnox environment. It isn't a model server, it isn't an AI tool, and it isn't explicitly AI-adjacent, but it is notionally useful if/when your agentic development workflows need credentials for APIs, package registries, cloud services, and so on. If you're copacetic sharing secrets with autonomous quasi-intelligent agents, you (or they) can give it a try by running flox activate -r flox/fnox.

This Has Been Radio. Free. Flox.

We’re Flox Labs and we’re solely responsible/liable for producing Radio Free Flox. Nominally we’re …

  • Rok Garbas, Flox engineer
  • Morgan Helton, Flox engineer
  • Brendan Magee, Flox engineer, drafted possibly against his will
  • Steve Swoyer, Flox probationary summer intern, aka “Esteban"
  • Various and sundry Flox PTBNL

...but we’re really anchored by the redoubtable Rok. Every week Rok tracks down software that either isn’t currently indexed by the Flox Catalog or doesn’t get refreshed as frequently as he’d like. When software’s absent from the Catalog, Rok uses Flox and Nix to build, package, and publish it reproducibly in CI. When it isn’t refreshed frequently enough, he sets up CI so new releases publish within hours.

Basically, Rok wants his Flox Claude Code environment to pick up the latest current-stable claude-code package as soon as it drops. In fact, Rok has an ongoing obsession with AI tools. He’s forever hunting down new/different CLI coding harnesses like DeepSeek TUI, agents like hermes, and anything/everything that supports them: proxies and routing tools, model servers, spec-driven dev kits, skill packs, agentic code-review tools, etc.

Rok also digs swine! He’s an avid participant in a Hungarian livestock share and is impatiently waiting for two shoats to mature. Out of a sense of delicacy, we won’t tell you what Rok plans to do once that happens.

Tune in next week to discover more incriminating facts about the people who bring you Flox.

Cue Radio Free Flox outro music...

Big man, pig man
Ha, ha, charade you are
You well heeled big wheel
Ha, ha, charade you are...

Keep learning with Flox

Interviews, tips, guides, industry best practices, and news.

View all posts