Quick start · How it works · autoloop · The catalog · Fluxor Cloud

An agent is only as good as the harness around it: what it knows about your project, the skills it can load, and the checks that stop it before it breaks something. Most teams keep that harness as a pile of prompt files copied between repositories, and each copy drifts on its own.

Fluxor keeps the harness as versioned items (rules, skills, hooks and files), assembles the right set for each project, and installs it into any repository with one command. Change an item once, and every repository that uses it gets the change on its next pull.

This repository is Fluxor Local: free and self-hosted, running on your machine with Docker. Fluxor Cloud keeps every machine and every teammate on the same harness.

Quick start

You need Docker and Python 3.

git clone https://github.com/CMolG/fluxor.git && cd fluxor
docker compose up -d --build
claude mcp add --transport http fluxor http://127.0.0.1:18191/api/fluxor/mcp

Ask your agent to "create a Fluxor harness project with the atoms and review-harness skills", or to run harness.bootstrap, which answers with the exact commands for your repository. By hand:

cd /path/to/your/repo && mkdir -p .harness/bin
curl -fsS http://127.0.0.1:18191/api/fluxor/v1/writer/fluxor-sync.py -o .harness/bin/fluxor-sync.py
python3 .harness/bin/fluxor-sync.py init --server http://127.0.0.1:18191/api --enroll local --project <project-id>

This is what that looks like (real output; machine name and paths shortened):

$ python3 .harness/bin/fluxor-sync.py init --server http://127.0.0.1:18191/api --enroll local --project 187ee517-... --storage local
fluxor-sync: enrolled as fluxor-sync on my-laptop on http://127.0.0.1:18191/api; the token is in ~/.config/fluxor/device-tokens.json
fluxor-sync: revision 1: 19 file(s) written, 0 removed, 12 downloaded

$ python3 .harness/bin/fluxor-sync.py status
project: 187ee517-...
server: http://127.0.0.1:18191/api
installed: revision 1, 18 file(s)
remote: up to date

From then on fluxor-sync pull brings the repository up to date, and fluxor-sync item publish turns a folder of yours into an item you can select into any project.

Hooks run code on your machine, so fluxor-sync never wires one without your consent (fluxor-sync approve). Everything listens on 127.0.0.1; set FLUXOR_LOCAL_TOKEN to require a token, which then also becomes the enrollment code.

How it works

flowchart LR
  subgraph fluxor ["Fluxor"]
    catalog["Catalog<br/>skills · roles · mods · hooks"]
    yours["Your own items"]
    project["Harness project<br/>the items one repo uses"]
    catalog --> project
    yours --> project
  end
  project -- "fluxor-sync pull" --> repo
  subgraph repo ["your-repo/"]
    harness[".harness/<br/>rules, skills, market"]
    claude[".claude/ · .codex/<br/>hook wiring"]
    agentsmd["AGENTS.md"]
  end
  repo --> agent(["Claude Code, Codex..."])

Loading

• Items, not copies. Every skill, hook and file is one item with a content hash. A project is a selection of items, and a pull is a snapshot of that selection: what changed is written, what was dropped is removed, and your local edits are kept, never overwritten.
• Atoms for each task. A deterministic engine (market/motor.py) mounts a role, its mods and its steps for the prompt at hand, so the agent gets the instructions this task calls for instead of all of them at once.
• One harness, every client. The same items wire into Claude Code and Codex, and the harness can live outside git when you want (--storage local).

autoloop: several agents, one backlog

autoloop is a skill in the catalog. It runs N agent sessions in parallel, each on a different card in its own git worktree, all integrating into one control branch, while a supervisor clears each session's context between cards.

flowchart TB
  backlog[("Officium backlog")]
  supervisor{{"autoloop supervisor"}}
  backlog -- "cards.next" --> supervisor
  supervisor --> s1["slot 1<br/>worktree · card A"]
  supervisor --> s2["slot 2<br/>worktree · card B"]
  supervisor --> s3["slot 3<br/>worktree · card C"]
  s1 --> control["control branch"]
  s2 --> control
  s3 --> control
  control -- "reviewed, integrated" --> backlog

Loading

It works the backlog of Officium, the board you and your agents share.

The catalog

The base catalog ships inside this repository (packages/fluxor-market) and is loaded into Fluxor Local on start: 49 roles, 58 mods, 11 steps and 13 flows for the atoms engine, 13 hooks, and these 15 skills:

Skill	What it does
atoms	Shows, explains or changes the role, mods and steps mounted for a task.
autoloop	Runs N agent sessions in parallel on the backlog, one worktree each.
backlog	A session protocol for working a tracker over MCP without agents stepping on each other.
review-harness	Checks that a harness is still wired and free of stale or broken references.
topology	Derives a workspace's real package graph and checks its layering rules.
reduction	A measure-first code-reduction campaign with 21 techniques for Java and React.
layout-discipline	Builds and reviews UI so nothing collapses, overlaps or overflows at any width.
ui-fluidity-verification	Frame, interaction and layout checks in real engines, with a statistical gate.
ux-psychology	Designs and audits interfaces with cognitive psychology, and flags dark patterns.
landing-page-marketing	Writes and audits landing pages that sell, checking every claim against facts.
traffic-to-funnel	Turns analytics data into a diagnosis and one funnel per product.
product-hunt-launch-readiness	Audits a product against Product Hunt's launch guide and builds the launch kit.
research-radar	Finds potentially novel techniques in a codebase and checks them against the state of the art.
portfolio-positioning	Turns a website into a hiring surface for its owner, page by page.
jira-agentic-requirements-pipeline	A two-phase requirements pipeline connected to Jira.

Local or Cloud

	Local	Cloud
Harness projects, items, fluxor-sync pull and publish	✓	✓
The base catalog and the atoms engine	✓	✓
MCP server for agents	✓	✓
autoloop and the Officium integration	✓	✓
Runs on your machine, nothing leaves it	✓
One harness shared by a team, across every machine		✓
The community marketplace, beyond the base catalog		✓
Personal packs from the CEET interview (your working judgment as agents and skills)		✓
The Fluxor web app: flow designer and harness library		✓
Nothing to host or back up		✓

Try Fluxor Cloud →

Questions

Is it free? Yes. Fluxor Local is free to use, modify and self-host for your own work and your company's, under the Elastic License 2.0. The one thing it does not allow is offering Fluxor to others as a hosted service.

Will it overwrite files I edited? No. A pull keeps your local edits and tells you which ones it kept; pull --force is the only way to take Fluxor's version over yours.

Does it run hooks without asking? No. A hook is wired only after you approve it, and a changed hook asks again.

Background

Fluxor carries the harness of FIERCE, a framework for software development where agents carry the implementation and a few experts hold the decisions:

Molina Galindo, C. (2026). FIERCE: Fast Implementation Ensuring Results by Capable Experts. doi:10.5281/zenodo.23972303

Repository layout

• server/: the API (Java, Quarkus): sync under /api/fluxor/v1, MCP under /api/fluxor/mcp. fluxor-sync.py ships inside it and is served at /api/fluxor/v1/writer/fluxor-sync.py.
• packages/fluxor-market/: the base catalog, its schemas and tools/marketplace.py.
• scripts/seed-catalog.mjs: regenerates the server's catalog seed from the catalog.
• scripts/smoke.sh: an end-to-end check, from docker compose up to a pulled harness.

This repository is published from the private monorepo where Fluxor is developed. Issues are welcome; pull requests are ported by hand.

License

Elastic License 2.0. See NOTICE.