FLORA

FLORA MCP Skills

Workflows for pointing your agent at FLORA and getting real work back — not just tool access.

FLORA's MCP server gives any agent the ability to generate images, video, audio and text, run saved multi-step techniques, and read and write a FLORA canvas. That is the raw capability. This repo is the judgment layer: when to use which model, what the gotchas are, what a good result looks like, and what to do when one comes back wrong.

Skills document tested workflows and measured observations. Model behavior and tool availability can change; check the connected server before following an example.

FLORA maintains these skills and publishes updates here through reviewed pull requests.

Install

One command. FLORA's MCP server is hosted, authenticated over OAuth, and needs no API key — the first call opens a browser to sign in.

Claude Code

claude mcp add --transport http flora https://agents.flora.ai/mcp

Codex

codex mcp add flora --transport http https://agents.flora.ai/mcp

Cursor, Windsurf, Zed, and other MCP clients — add to your MCP config:

{  "mcpServers": {    "flora": {      "type": "http",      "url": "https://agents.flora.ai/mcp"    }  }}

Claude.ai (web and desktop) — add it as a custom connector:

1. Copy https://agents.flora.ai/mcp
2. Go to Settings → Connectors → Add custom connector (direct link)
3. Paste, sign in, and start generating

ChatGPT — FLORA is in the app directory; enable it there.

Then add the skills

The server is the capability; the skills are how your agent knows what to do with it.

Claude Code — one command, skills and server together. This repo is also a plugin marketplace, so installing it wires up the MCP server and every skill:

claude plugin marketplace add florafauna-ai/flora-mcp-skillsclaude plugin install flora-mcp-skills@flora

Any other agent — copy the skills in:

git clone https://github.com/florafauna-ai/flora-mcp-skills.gitcp -r flora-mcp-skills/skills/* .codex/skills/       # Codexcp -r flora-mcp-skills/skills/* .cursor/rules/       # Cursor

Claude.ai — there is no plugin install, so pick one of three. claude plugin is Claude Code only; on web and desktop the skills go in by hand.

A Project, if you use FLORA regularly. Create a Project, open Set custom instructions, and paste in the contents of the skill you want — for example skills/flora-batch-generate/SKILL.md. Every chat inside that Project then has it, with nothing to re-paste. This is the most reliable option and it costs nothing.

Upload it as a Skill, for something reusable across Projects. Claude.ai takes a zipped skill folder, and this repo's layout is already the right shape — one directory per skill with a SKILL.md inside it. Zip the one you want:

git clone https://github.com/florafauna-ai/flora-mcp-skills.gitcd flora-mcp-skills/skills && zip -r flora-batch-generate.zip flora-batch-generate

Then upload it from Claude's settings, under skills/capabilities. Zipping from the repo rather than downloading a prebuilt archive means you always get the current version.

Or just ask for it inline — see the raw-URL option below, which needs no setup at all and works in a normal chat.

Or no install at all — point your agent straight at one skill:

Read https://raw.githubusercontent.com/florafauna-ai/flora-mcp-skills/main/skills/flora-batch-generate/SKILL.mdand follow it for these 5 products:
  PO-01  a matte black ceramic pour-over coffee dripper  WB-04  a brushed steel insulated water bottle  AP-09  a folded oatmeal linen apron  CB-02  a walnut wood cutting board, rectangular  SB-03  a stack of three speckled stoneware bowls
One consistent studio style across all five: seamless warm-grey backdrop,soft softbox from camera left, subtle contact shadow, 85mm, centred, no props.

Keeping the skills up to date

New skills get added to this repo, and existing ones get corrected as we measure things that turn out to be wrong. None of that reaches you on its own. An installed plugin is a frozen snapshot — it will not tell you a new skill exists, and it will keep working happily on an old copy. Updating is something you do, not something that happens.

Claude Code — two commands, no reinstall and no re-auth. Your MCP connection is untouched:

claude plugin marketplace update floraclaude plugin update flora-mcp-skills

The first refreshes this repo's marketplace listing; the second pulls the new version. Skipping the first means the second has nothing new to find. Changes apply on your next session — plugins load at start-up, so restart your agent afterwards.

If it reports already at the latest version, you are genuinely current. claude plugin list shows the version you are on.

Any other agent — re-copy. There is no update command, so pull the repo again and overwrite:

cd flora-mcp-skills && git pullcp -r skills/* .codex/skills/        # Codexcp -r skills/* .cursor/rules/        # Cursor

Claude.ai — re-paste or re-zip. A Project's custom instructions and an uploaded Skill are both snapshots taken when you added them. Paste the current SKILL.md in again, or re-zip the folder from a freshly pulled clone and re-upload.

Raw URL — nothing to do. That path always serves main, so it is current the moment we push. If you want to be sure you are reading today's version of a skill, use it.

Start here: stop chatting, start batching

The single highest-leverage thing you can do with FLORA's MCP server is stop treating it as a chatbot and start treating it as a pipeline.

One request at a time, waiting for each, is the default an agent falls into and it is the slowest possible use of the server. FLORA runs generations concurrently. The serialisation is entirely on the agent's side.

Measured on a real six-item batch: fired together, 15.8 seconds. The same six run one after another: about 150 seconds. Identical cost, identical outputs, one tenth the wall clock — and the gap widens linearly with the size of your list.

flora-batch-generate is the flagship skill in this repo. If you read one thing here, read that. It turns a product list, a set of campaign variants, a folder of source images, or a spreadsheet into one parallel batch, polls it correctly, and tells you exactly what failed and why.

Skills by use case

Scale — many items, one treatment

Refine — get one thing right, without burning money

Composite — animate what already exists

Sequence — script to moving pictures

Transform — you already have the image

Reformat — one asset, every placement

Present — turn finished work into a deliverable

Ship — all the way to the ad platform

Watch — keep a team plugged in

Check — judge it before it ships

Iterate — build on what exists

★ = flagship, most heavily tested.

Try it

Copy one of these. Skills fire on what you ask for, not on being named — you should never have to type a skill's name.

flora-batch-generate

Here are 24 products from our spring catalogue. Generate a PDP shot for each one — same studio setup on every product, seamless warm-grey, soft key from camera left. Total the cost before you spend anything.

Take this campaign line and give me 12 localised versions, one per market, same art direction throughout. Markets are in the attached sheet.

I have a folder of 40 product URLs. One hero image each, consistent lighting. Tell me what it'll cost first.

flora-refine-loop

Generate a single white ceramic mug on a plain black surface. Handle pointing right, top-down bird's-eye view looking straight down, no other objects. Keep refining until it matches — max 3 rounds.

That's not quite right, try again.

(The skill should refuse to re-roll blind and ask what "right" means. If it just regenerates, the skill didn't fire — that's the bug to look for.)

A photo of a clock face showing exactly 4:37. Refine until the hands are correct.

(Models are bad at clock hands. Correct behaviour is stopping once the same defect survives a real prompt change, and handing you options — not burning more rounds.)

flora-script-to-video

Turn this script into a video. Same character and lighting across every shot.

INT. DINER — NIGHT. Rain on the window. A woman sits alone with a coffee.She checks her watch. The door opens. She looks up.

Storyboard this scene and animate it — 6 shots, 5 seconds each, 16:9. Show me the keyframes before you spend anything on video.

Here's a 30-second ad script and our brand character sheet. Build the keyframes off that reference, then animate. Quote the video stage separately.

flora-mockup-deck

Mock this poster up out of home — I want to see it on a wall, a bus shelter, that kind of thing. Plus social sizes.

What would this campaign key visual look like in the wild? Square-on shots, and don't touch the artwork.

Give me a placement deck for this ad, with 1:1, 4:5 and 9:16 resizes for social.

Create a campaign deck of a flower.

(No artwork attached, so the skill writes the poster first — one generation — shows it to you, and only places it once you say yes. Four wrong placements is the failure it is avoiding, and they are not re-rollable.)

flora-brand-ad-pipeline

Here's the strawberry launch — six SKUs, our usual treatment. Get them staged for me to review.

Build the ad set for the new store opening: generate, archive, host, and leave the Meta creatives paused.

Re-run the summer campaign — two of the ads didn't make it into Meta last time.

(The manifest is the point: a re-run should repair the two missing rows, not regenerate and re-upload all of them.)

flora-pdp-deck

Here's a photo of our candle. I need PDP shots for the product page, and a deck at the end.

Product shot and a full-length model frame attached. Shoot it on her, studio background, and de-brand it.

I've only got this one product photo. What can you actually get me from it?

(The skill should say plainly that one photo buys one plate, not four — asking a single reference for four angles returns four near-duplicates and bills for all of them. If it promises a full multi-angle set, that's the bug.)

flora-run-technique

Put this product on a clean white background.

Match the lighting in this shot to the reference image.

What can FLORA actually do to an image I already have?

flora-canvas-iterate

What's in my Meridian project?

Pull up the hero images I made last week and redo them warmer — less blue in the shadows.

Show me what I've generated in this project and what it cost.

What a good skill looks like here

The bar is set by flora-batch-generate and flora-mockup-deck. Both share a shape, and it is not "a description of the tools":

• A law. One sentence naming the invariant the skill exists to protect. "N items is one wait, not N waits." "A re-run with an unchanged prompt is not a refinement." If you cannot write that sentence, the skill is not ready.
• Measured numbers. 265ms, 3.45x, 17% undercount, limit: 100. Not "fast", not "expensive", not "usually works".
• Named gotchas. The specific thing that silently ate someone's budget, written down with what it looked like when it happened. The most valuable content in this repo is the failure someone already paid for.
• What it does not do. Where the skill stops, stated plainly, so an agent does not improvise past its tested edge.

Thin wrappers around tool descriptions are not useful — agents can already read tool schemas. The value is the judgment that only comes from running it.

Contributing

Some of the best FLORA workflows are already being built by users, independently of us. Two we know of:

• Pedro Padilla's agentic marketing pipeline — an end-to-end campaign flow driven through the MCP server rather than the canvas.
• Will Bovill's Figma + FLORA storyboard workflow — pulling frames from Figma, generating against them, and assembling the result back into a board.

Neither started as a feature request. Both are the kind of thing this repo exists to surface, so that the next person does not have to rediscover it.

If you have built a workflow that works, send it back. The bar is the four bullets above. Open an issue describing the workflow or correction; a maintainer will review it and include accepted changes in a future update.

Reference

• FLORA MCP endpoint — https://agents.flora.ai/mcp (streamable HTTP, OAuth)
• Workflow instructions — flora_discover_skills lists available skills; pass a skill name to load its instructions.
• API and SDK docs — docs.flora.ai.
• Calling FLORA from a skill — use the named tools: flora_create_generations, flora_list_generations, flora_run_technique, flora_list_models and friends. Create 1–20 generations with { "generations": [...] }, including a one-item array for a single generation. Retain successful run IDs and poll with { "run_ids": [...] }; include technique_id for technique runs. Omit run IDs only to list history. Never retry a whole batch because one submission failed. The hosted server no longer has an execute code tool; flora_list_tools lists what the connection offers.

Connection reference

Check the connected server's advertised tool list and schemas before running a workflow. The examples target the plural hosted generation tools; older cached catalogs need refreshing.

endpoint      https://agents.flora.ai/mcp   (streamable HTTP, Clerk OAuth)API base      https://app.flora.ai/api/v1SDK methods   generations.create · generations.list · generations.retrieve              models.list · projects.create · assets.create              techniques.runs.create · techniques.runs.retrieve

Two things that do not exist and are easy to reach for: models.retrieve (use models.list().find()) and runs.action (use the flora_run_action tool).

FLORA also ships an in-browser MCP surface (WebMCP) that operates a live canvas directly. It has a different, larger tool vocabulary — flora_run_batch, flora_get_batch_status, flora_get_generations, flora_edit_layers and others — and some skills explicitly distinguish that path. If you are writing a skill against the browser surface, check its tool names separately; they are not interchangeable with the hosted server's.

Data and privacy

This plugin is Markdown skills plus a reference to FLORA's hosted MCP server. It has no hooks, no background processes and no telemetry of its own.

• What goes to FLORA. When Claude calls a FLORA tool, the plugin sends what that call needs to https://agents.flora.ai/mcp: prompts, reference images and files you provide, model and technique choices, and project or canvas content you ask it to read or change. You sign in with your own FLORA account over OAuth; nothing is sent until you connect. Generations run against your FLORA account and bill its usage, and results are stored in your FLORA workspace and served from media.flora.ai.
• What stays local. flora-motion-compositor includes one readable Python script, scripts/comp.py, that renders frames on your machine with ffmpeg and, for SVG input, headless Chrome. It makes no network calls. The deck skills render HTML and PDFs locally with headless Chrome in the same way.
• Other services, only when you ask. A few skills describe steps in tools you already use — fetching trend sources you name (flora-trend-board), posting boards to a team channel, or leaving paused ads in Meta Ads Manager (flora-brand-ad-pipeline). Those steps run through connectors or credentials you set up yourself; the plugin ships none.

FLORA's Privacy Notice and Terms of Service apply to data sent to FLORA. Questions: open an issue or email [email protected].

Licence

MIT. See LICENSE.