Alchemy MCP server

Alchemy

Alchemy is a drop-in ruleset that stops your AI coding agent from writing like an LLM: no em-dash pile-ups, no "not just X, it's Y", no stray delve. You add it once, your agent reads it before it writes, and the READMEs, commits, and docs it produces start sounding like a person wrote them.

Turn lead into gold. It's one file, about 1000 tokens, that lives next to your agent's other instructions: ALCHEMY.md. Everything else here just gets it into your project.

How do I stop my AI from writing like an LLM?

Install it in your project and point your agent at it. Pick whichever fits your stack:

# Node / npm
npx @fernforge/alchemy init

# Python / pip
pip install alchemy-writing
alchemy init

init writes ALCHEMY.md into your project and links it from whatever agent config you already have: CLAUDE.md, AGENTS.md, .cursorrules, or Copilot instructions. From then on the agent reads the rules as part of its normal context. To print the ruleset to stdout without installing anything:

npx @fernforge/alchemy print   # or: alchemy print

Prefer to do it by hand? Copy ALCHEMY.md into your repo and reference it from your agent's instruction file. That's the whole mechanism. No service, no API, no account.

Before and after

Here's the kind of sentence Alchemy catches. Same fact, two ways of saying it:

Before

Redis isn't just a database — it's a powerful, robust caching layer that unlocks the full potential of your entire stack.

After

Redis is an in-memory store. Put your cache and sessions in it and your database stops being the bottleneck.

The "before" stacks four tells in one sentence: the "isn't just X, it's Y" non-contrast, the adjective pile (powerful, robust), the lone dramatic em-dash, and "unlocks the full potential". None of it says what Redis does. The "after" drops all of it and states one concrete thing you can act on. That's the whole move, applied everywhere the agent writes.

Which LLM tells does it remove?

The concrete ones, by name:

• Empty contrasts: "not just X, it's Y", "it's not about A, it's about B".
• Staccato trios and rule-of-three padding: "fast, reliable, scalable", "No X. No Y. Just Z."
• Stacked adjectives standing in for facts: robust, seamless, powerful, comprehensive.
• Em-dash pile-ups. The real tell is density, so the rule is: use them rarely.
• Filler vocabulary: delve, leverage, harness, unlock, elevate, tapestry, realm, showcase.
• Throat-clearing openers ("In today's fast-paced world", "At its core") and recap closers ("In conclusion", "Ultimately").
• Restating the prompt before answering it.

The fix is never a synonym swap. It's replacing the vague word with a specific fact: not "robust" but "handles 10k req/sec". The rules push the agent toward the fact.

The principle underneath: no single word proves a machine wrote something. The signal is many of these clustered together over flat, evenly-weighted text. Alchemy targets the clustering, so it thins the tells without stripping a word that's genuinely the right one.

Use it with Claude, Cursor, or Copilot

init detects the config files already in your project and links the ruleset from them, so the same ALCHEMY.md works across agents:

• Claude Code / Claude reads it via CLAUDE.md.
• Cursor reads it via .cursorrules.
• Copilot reads it via its instructions file.
• Any agent that supports AGENTS.md picks it up there.

One ruleset, whichever agent you're driving.

Use it as an MCP server

To serve the rules on demand across every project instead of committing a file to each repo, run Alchemy as an MCP server. Add it to your client config (Claude Desktop, Cursor, Cline):

{
  "mcpServers": {
    "alchemy": { "command": "npx", "args": ["-y", "@fernforge/alchemy-mcp"] }
  }
}

Shortened. The full README is on GitHub.