LEARN / AGENTS AND SKILLS

Agents and skills

Flox does not sell an agent. It sells the ground under one: a pinned, hash-addressed environment that a human or a coding agent activates the same way. Flox ships a skills library (flox, floxify, flox-debug), publishes explicit instructions for AI coding assistants in its llms.txt, and, as of an October 2026 retirement, no longer maintains its own MCP server.

This page is drawn from https://flox.dev/llms.txt, https://github.com/flox/flox-skills, https://flox.dev/deterministic-agents/, https://flox.dev/blog/a-turnkey-toolkit-for-agentic-development-with-flox/, https://flox.dev/blog/next-level-agentic-development-with-flox/, https://flox.dev/blog/focusing-on-the-code-not-the-cruft/, https://flox.dev/blog/byoa-needs-a-floor-that-floor-is-flox/, https://flox.dev/blog/code-and-its-runtime-should-travel-together/, https://flox.dev/docs/man/flox-activate and https://flox.dev/docs/changelog/2025.

The pitch in one paragraph

The deterministic-agents page says it plainly: “The harness is probabilistic by design. The environment under it should not be.” (https://flox.dev/deterministic-agents/). One manifest is pinned to the same toolchain under every harness, on a laptop and in the cloud (https://flox.dev/deterministic-agents/). Flox owns one layer, the environment and its activation hooks, and the page says the ecosystem owns “Harness and sandbox internals”, meaning how Claude Code reasons, how an MCP server talks to GitHub, and how a devcontainer or E2B box isolates a run (https://flox.dev/deterministic-agents/).

Say: “Flox does not replace your agent or your sandbox. It makes what is inside them identical every time.” Ask: “When your agent fails, can you tell a real failure from a setup failure?” The page argues that drift between laptop, sandbox and CI forces the agent to debug “the substrate” instead of doing the work (https://flox.dev/deterministic-agents/).

flowchart TB
  H[Agent harness<br/>Claude Code, Codex, others] --> S[Sandbox<br/>devcontainer, Docker, hosted]
  S --> E[Flox environment<br/>manifest.toml + lock]
  K[Skills<br/>flox, floxify, flox-debug] --> H
  E --> P[Pinned packages<br/>and services]
  E --> F[FloxHub<br/>generations]

Whiteboard diagram (rendered from the source mermaid)

Deterministic agents: what is claimed

Anatomy. The page lays out a stack of harness, sandbox, plugins and MCP servers, skills, and the Flox environment at the bottom; “Swap the harness, change sandboxes, add an MCP server: the base layer does not move.” (https://flox.dev/deterministic-agents/)
Runs under the tools you have. It composes with VS Code, Docker and devcontainers: open the repo, attach a devcontainer or Docker box, then run flox activate (https://flox.dev/deterministic-agents/).
Remote sandboxes. For Coder, DevPod, Daytona, E2B, Fly or Cloudflare, the page says the manifest resolves to the same closure it did on the laptop, and an activation hook can fetch short-lived credentials so nothing is written into a manifest, image or repo (https://flox.dev/deterministic-agents/). Treat the per-vendor claims as marketing copy until a customer has run them [verify per sandbox vendor].
One command. The page’s own example is flox activate -- claude after pinning Python, Node and PostgreSQL in the manifest (https://flox.dev/deterministic-agents/).

The flox-skills library

Flox publishes skills that teach AI coding agents how to use Flox properly (https://github.com/flox/flox-skills). The README says they work with Claude Code, Codex and any agent supporting the skills.sh standard (https://github.com/flox/flox-skills).

Skill	What it does	Source
`flox`	Creates and manages environments: installs and pins packages, sets up services, builds and containerizes applications, publishes to FloxHub, composes environments	https://github.com/flox/flox-skills
`floxify`	Onboards an existing repo: detects runtimes, services and build tools, then writes `.flox/env/manifest.toml` so `flox activate` is the only setup command; converts DevBox, Dev Containers, Brewfiles and asdf or mise pins	https://github.com/flox/flox-skills
`flox-debug`	Diagnoses catalog resolution, such as a package that will not resolve, `flox install` picking an old build, or a “constraints too tight” failure	https://github.com/flox/flox-skills

Install paths in the README (https://github.com/flox/flox-skills):

# Claude Code
claude plugin marketplace add flox/flox-skills
claude plugin install flox@flox-skills

# Codex
codex plugin marketplace add flox/flox-skills
codex plugin add flox@flox-skills

# Other agents (skills.sh, requires Node.js)
npx skills add flox/flox-skills

Two more facts from the README: the plugin is also packaged in an agent-stacks catalog (agent-stacks/agent-plugin-flox), and the older flox/skills-flox package is no longer published (https://github.com/flox/flox-skills). The README also says the skills are evaluated against a no-skill baseline, with methodology under evals/ (https://github.com/flox/flox-skills); I did not read the eval results, so quote no numbers [verify].

Evidence from Flox’s own writing: a developer advocate tried flox and floxify with Claude Code on an Apple Silicon Mac, and reports a working FastAPI embedding service from flox, and a six-service Mastodon development stack from floxify, “warts and all” (https://flox.dev/blog/focusing-on-the-code-not-the-cruft/, published 2026-10-01). The floxify flow checks package names against the catalog before writing the manifest, then runs a gate that tests resolution, hooks and activation before it signs off (https://flox.dev/blog/focusing-on-the-code-not-the-cruft/). That is one author’s trial on one machine, so frame it as an example, not a benchmark.

The October 29, 2026 post makes the case for floxify: README files are not an authoritative definition of what a repo needs, so commit a Flox environment, and the agent reverse-engineers the repo once, up front (https://flox.dev/blog/code-and-its-runtime-should-travel-together/).

llms.txt: the instructions for AI coding assistants

The site index at https://flox.dev/llms.txt (last updated 2026-07-27 per the file) carries an “Instructions for AI coding assistants” block. These lines are the ones to know, quoted exactly:

The command to enter an environment is flox activate. It is not nix develop or nix-shell. Flox is built on Nix but has its own CLI surface.

Do not run bare flox activate in a non-interactive harness. With no command argument it launches an interactive sub-shell and will hang. Use flox activate -d <dir> -c '<command>', which runs the command with hooks and profile scripts applied, or flox activate -d <dir> -- <command>, which is exec mode and does not run profile scripts.

A Flox environment is defined by .flox/env/manifest.toml. Do not invent manifest fields. Check the manifest reference at https://flox.dev/docs/man/manifest.toml.md before writing one.

Package names come from the Flox Catalog and follow catalog pkg-paths, not nixpkgs attribute paths in every case. Verify with flox search <term> or flox show <package> rather than guessing.

(all from https://flox.dev/llms.txt). The same block says Flox runs on Linux and macOS with “no native Windows build; Windows users run it under WSL2”, and points at the skills install commands (https://flox.dev/llms.txt).

Safe non-interactive usage

The activate man page defines the modes (https://flox.dev/docs/man/flox-activate):

Invocation	Behavior
`flox activate` in a terminal	Interactive sub-shell
`flox activate -c CMD`	Runs CMD in the environment as if in an interactive shell; shell features like `&&` work
`flox activate -- CMD`	Execs CMD directly after activation, except `[profile]` scripts
`eval "$(flox activate)"`	In-place: produces commands for the parent shell to source

For an agent or script, use flox activate -d <dir> -c '<command>' so it cannot hang on a shell prompt (https://flox.dev/llms.txt). Two more rules from the docs: activating from a cached lockfile “does not require a FloxHub account” while a private remote environment (-r) does (https://flox.dev/docs/man/flox-activate), and activating a remote environment runs a shell hook that “may execute arbitrary code”, which is why --trust exists (https://flox.dev/docs/man/flox-activate). Advice: never pass --trust for an environment you have not read.

Since October 20, 2026, an agent that has to resolve new packages needs a FloxHub login, while an environment with a valid committed lockfile activates without one (https://flox.dev/blog/flox-catalog-login-requirement-what-you-need-to-know/). Offline agent sandboxes still need the packages available locally, because tokens “cannot substitute for network connectivity” (same post).

The agentic posts, one line each

Post	Date	Takeaway
Agentic Development Environments with Nix and Flox	2025-09-26	Agents use `flox search`, `flox show` and `flox install`; a Flox-aware CLAUDE.md, AGENT.md or FLOX.md gives them context (https://flox.dev/blog/next-level-agentic-development-with-flox/)
BYOA Needs a Floor. That Floor Is Flox.	2026-03-18	Governance layers wrap a runtime whose contents are “assumed”; Flox is pitched as the pinned floor (https://flox.dev/blog/byoa-needs-a-floor-that-floor-is-flox/)
Focusing on the Code, Not the Cruft	2026-10-01	Hands-on trial of the `flox` and `floxify` skills (https://flox.dev/blog/focusing-on-the-code-not-the-cruft/)
Code and Its Runtime Should Travel Together	2026-10-29	Commit the environment with the code; `floxify` creates it for existing repos (https://flox.dev/blog/code-and-its-runtime-should-travel-together/)
A Turnkey Toolkit for Agentic Development with Flox	2026-11-04	“Nearly 80” agent environments in one repo; each folder is its own environment (https://flox.dev/blog/a-turnkey-toolkit-for-agentic-development-with-flox/)

The turnkey repo is github.com/flox/agentic-development-with-flox. None of its environments commits a manifest.lock, so first activation needs a free FloxHub login to resolve through the catalog, and later activation with a saved lockfile does not (https://flox.dev/blog/a-turnkey-toolkit-for-agentic-development-with-flox/). The same post says Flox sponsors Agent Stacks, whose specification “is still a draft”, with sandboxing and secrets management still “works in progress” (https://flox.dev/blog/a-turnkey-toolkit-for-agentic-development-with-flox/).

MCP: current status, stated exactly

The sources disagree because they were written at different times, so say the date with the claim.

Current (retired). The turnkey post dated 2026-11-04 says: “Flox retired its own MCP server in October 2026”. The reason it gives is that, as a thin wrapper around the CLI, agents could use the CLI directly, and Flox “focused its efforts on skills” (https://flox.dev/blog/a-turnkey-toolkit-for-agentic-development-with-flox/). The post gives no day within October, and I found no separate retirement notice [verify the exact date and any migration guidance with Flox].
Cleanup in the repo. The same post says the claude-code environment’s activation hook removes an old Flox MCP registration if it finds one, and the codex environment removes the old entry (https://flox.dev/blog/a-turnkey-toolkit-for-agentic-development-with-flox/).
Earlier history. The 2025 changelog lists “Flox Agentic MCP server” on December 17 (https://flox.dev/docs/changelog/2025). A March 2026 post still describes it as live: “The Flox MCP server lets any MCP-capable coding tool” author, build and publish environments (https://flox.dev/blog/byoa-needs-a-floor-that-floor-is-flox/, 2026-03-18).
Stale references. Posts from December 2025 to March 2026 still tell readers to install the Flox MCP server (https://flox.dev/blog/reproducible-development-environment-in-any-repo-part-1/, https://flox.dev/blog/gpu-optimized-pytorch-builds-made-easy-with-flox-and-nix/). Do not repeat that advice to a customer; point them to skills plus the CLI.
MCP servers as workloads still work. Flox environments can define MCP servers as services (https://flox.dev/blog/a-turnkey-toolkit-for-agentic-development-with-flox/), and the May 2026 roundup mentions environments such as flox/serena that bundle an MCP server (https://flox.dev/blog/radio-free-flox-for-may-23rd-2026/). Retiring Flox’s own MCP server does not stop you pinning other MCP servers in a manifest.

Say: “Flox moved from an MCP wrapper to skills over the CLI in October 2026. If a blog post tells you to install the Flox MCP server, it predates that.” Ask: “Which agent do you run, and is it on the skills install path above?”

← Containers and Kubernetes · FloxHub, orgs, plans, compliance →