<!-- mcp-name: io.github.RNVizion/rnv-color-mcp -->

RNV Color MCP

A remote MCP server for a complete color workflow: mix, convert, harmonize, and remember palettes, called in plain language by Claude (or any MCP client), and by anything else that speaks MCP.

Why this exists

The color logic already lived in my desktop suite: a mixer, a palette manager, a picker. Instead of rebuilding it for every new project, I lifted the engine out once and exposed it as a single server. A Claude conversation calls it today; a fashion design app will call the same backend tomorrow. Build the engine once, let both consume it.

Underneath, it's a small thesis about working with LLMs: a model is great at deciding what you want and terrible at exact arithmetic. So the model picks the tool and the intent, and the tool owns the precise values. The server resolves or it refuses; it never guesses a color.

What it does

Every color input accepts a hex (#d2bc93), a CSS name (red), an RNV brand name (brand gold, near-black), or a saved-palette reference (Spring line, or Spring line:2 for its second swatch). Brand names win over CSS names on collision (gold, blue, teal); css:blue forces the universal one.

Connect in 30 seconds

This is a hosted server, so there's nothing to install. In Claude: Settings → Connectors → Add custom connector, then paste:

https://rnvizion-rnv-color-mcp.hf.space/mcp

Leave auth blank, add it, then toggle it on in a chat with the + menu.

Try it

Once connected, just talk:

"Save a palette named Spring line: near-black and brand gold." "Pull my Spring line palette and give me three complementary accents for outerwear." "Mix crimson and royalblue like real pigment."

The first call saves; the second composes get_palette → generate_harmony; the third runs the Kubelka-Munk paint model, so the blend darkens the way mixed pigment actually does, not the way averaged light does.

Authentication

The server implements OAuth 2.1 resource-server authentication with enforced per-tool scopes; the public endpoint runs with it disabled, so connecting by URL works with no setup.

When enabled (RNV_AUTH=1 plus a key source), the server validates bearer tokens against issuer, audience, expiry, and signature; serves RFC 9728 protected resource metadata at /.well-known/oauth-protected-resource/mcp; returns a spec-compliant WWW-Authenticate challenge on 401; and enforces two scopes: read covers every read-only tool, write covers save_palette, the only tool that mutates anything. A token without the write scope does not see save_palette in its tool list at all; out-of-scope tools are hidden rather than refused, so nothing leaks about what exists behind a scope you lack.

Enforcement is covered by the test suite, run in CI on every push (status badge at the top of this file), spanning the token-validation matrix (missing, malformed, wrong issuer, wrong audience, expired, valid) and end-to-end scope enforcement over real HTTP. The suite mints its own keys and requires no credentials. Beyond auth, it pins parts of the engine: the brand vocabulary mirror; the contrast, difference and placement contract; selector and input refusals; color-name resolution; what the mix modes compute; the hsv and hsl conversions; the palette write path; store durability. Harmony and text outputs are exercised by the smoke scripts and are not yet under regression test.

Moving from the self-issued development key to a real identity provider is configuration, not code: point RNV_AUTH_JWKS_URI at the provider's JWKS endpoint and set the issuer and audience to match.

With RNV_AUTH on and no key source set, the server refuses to start rather than coming up unprotected.

Run it yourself

pip install -r requirements.txt
python server.py                  # Streamable HTTP on $PORT (default 7860)

pip install -r tests/requirements-dev.txt
python -m pytest                  # the suite; needs no credentials
python tests/server_test.py       # smoke: exercises every tool in-process

Set HF_TOKEN to write palettes through to a private Hugging Face Dataset for durable storage. The suite and both smoke scripts ignore HF_TOKEN and use a throwaway store, so they can never write to a real Dataset.

Running a copy?

The code is MIT-licensed; you're free to rebuild, host, and modify it, and I'd rather you did than didn't. What a license can't carry is the environment.

Some capabilities depend on how a deployment is configured, not on the code alone:

• Palette persistence needs HF_TOKEN and a writable Dataset. Without it, saves land in the

container's local working copy and vanish with it; save_palette returns durable: false with a durable_reason naming the step that stopped it.

• Scoped authorization needs RNV_AUTH and a configured issuer. Unset, the server runs open —

which is correct for a public demo and wrong for anything else.

• Anything that fetches needs outbound network access.

A rebuild missing one of those usually doesn't fail loudly; it returns success and drops the result. That is worse than an outage, because an outage tells you.

So: if you're running a copy and something behaves oddly, check the environment before you check the code, and check against the canonical deployment before you file anything.

Canonical endpoint: https://rnvizion-rnv-color-mcp.hf.space/mcp Canonical source: https://github.com/RNVizion/rnv-color-mcp

A copy served from another URL may be older, modified, or differently configured. That's fine and allowed; it just isn't this.

The tools resolve or refuse; they don't guess. A copy can only keep that promise if the ground under it was checked first.

Notes

• Brand colors are mirrored, not owned. The vocabulary lives in

engine/brand_vocab.py, mirrored from engine/brand.py in RNVizion/rnv-brand and corrected when drift is detected against it. It is carried locally on purpose: resolve_color is the hot path, and a fetch there would have to answer what happens when it fails — fail closed and the server refuses every color, fall back and the local copy is needed anyway, guess and the promise above is already broken. Identifiers match upstream's since the register retired local names on 2026-08-17; the check compares values. A scheduled job compares the mirror with upstream daily and gates nothing. The register and its reasoning live in BRAND_COLORS.md, in rnv-brand.

• Engine is dependency-free. The color math, harmony, and text logic are pure standard

library, lifted Qt-free from the desktop apps. Only the server layer needs fastmcp.

• Honest by design. An unknown color name is refused, not guessed. An unverifiable token is

refused, with a reason. Same principle at both layers.

Stack

Python · FastMCP (Streamable HTTP) · Hugging Face Spaces (Docker) · huggingface_hub for durable palette storage.

Built by Christian "RNVizion" Smith.