mcode

☿ The name

Mercury · diffusion · code.

MercuryCode is the project name. mcode is the command you type. Behind both runs Mercury 2 — Inception Labs' diffusion LLM, named after the planet that races around the sun in 88 days.

Roman messenger, alchemical quicksilver, the fastest planet, and now: the fastest loop on your terminal.

Why diffusion changes the loop.

Most LLMs write code one token at a time. Diffusion writes a whole region at once and refines it. The practical effect is that single-file edits — rename across file, add JSDoc, restructure, idiomatic refactor — return in roughly 300 milliseconds instead of seconds.

mcode leans on this. The agent picks edit_with_ai for structural changes and reaches for edit_file only when an exact-string replacement is intended. The result is a tighter feedback loop and far less waiting on the model.

Four-frame diagram showing diffusion denoising from t=1000 noise to t=0 resolved code.

A typical session.

$ mcode -y "rename multiply to mul in src/index.js, add JSDoc"

⏵ AIEdit src/index.js
✓ aiedit
  /src/index.js  +24 -8

⏵ Run    $ npm run build
│ > tsc
│ BUILD_OK
✓ run  exit 0

● Renamed multiply → mul (3 sites). JSDoc added to both functions.
  Build verified.

  elapsed: 2.6s

From prompt to verified build in under 3 seconds.

What it does.

Agent that picks the right tool

mcode reads files, runs commands, edits in place, and recovers. It picks edit_with_ai for structural changes and edit_file for exact replacements — without ceremony.

mcode -y "extract a helper from this duplicated logic"

A marketplace, built in

Browse and install plugins from any Claude Code-compatible marketplace. Skills, hooks, MCP servers — they all just work.

mcode plugin marketplace add wshobson/agents
mcode plugin install code-reviewer@wshobson

Speak MCP fluently

Brave Search, GitHub, filesystem — any MCP server connects via stdio. The agent gets the tools, you get an audit trail.

{ "mcpServers": { "brave-search": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-brave-search"] } } }

Quiet by default.

Writes and shell commands ask for approval. --yolo only when you want it.
/undo reverts the last edit. Save you from a bad LLM moment.
Post-edit syntax check on .ts/.js/.json/.py — broken code surfaces immediately.
Project-aware hooks: mcode init writes a hooks.json that auto-runs your npm test after edits.
Read-only mode disables every write tool.

Install in 30 seconds.

git clone https://github.com/south0120/MercuryCode.git
cd MercuryCode/mcode
npm install
npm run build
npm link

mcode is now on your $PATH. npm distribution is coming once the API stabilizes.

Configure

export INCEPTION_API_KEY=sk_...
mcode

Need a key? Sign up at Inception Labs. Mercury 2 costs roughly $1 per million tokens — about a tenth of frontier-class chat models.