Lucicoder

Approval modes

The approval mode decides how far Lucicoder goes before it asks you. Switch at any moment — with /mode, for example /mode ask.

• Chat answers and explains. It touches nothing.
• Ask asks before every tool it wants to use.
• Agent reads and searches freely, and asks before it edits a file or runs a command.
• Auto works on its own — still inside the command denylist, still with a checkpoint for every turn.

Preview is a fifth way to work: edits are staged for your review instead of written. It is described below.

Diffs, checkpoints and undo

Every edit Lucicoder proposes is shown as a red-and-green diff before it is written.

Every turn is a checkpoint. One click reverts everything that turn changed, across every file it touched — so trying something is cheap.

For anything older, Local History keeps every save and every Lucicoder edit as a version outside the repository: a timeline, side-by-side diffs, labels and one-click restore. Open it from the command palette.

Preview mode

In Preview mode, Lucicoder's edits are staged instead of written. You review the multi-file diff — and can edit it — then apply or discard each file. A file that changed on disk after its edit was staged is flagged, so nothing is overwritten by surprise.

Two refactoring commands use the same review:

• Rename Symbol with Preview (Ctrl+F2) stages the language server's rename.
• Refactor with Lucicoder (preview) (Ctrl+Shift+R) drafts the request from your selection.

Commands and background tasks

A command Lucicoder runs is stopped after 120 seconds. Dev servers and watchers run as background tasks instead — Lucicoder starts, reads and stops them with run_background, task_output and task_stop — and /tasks lists them.

A hard denylist blocks dangerous commands in every mode, whatever the prompt says. Commands you trust can go on the project's “always allow” list.

Slash commands

Type / in the Lucicoder panel to see them all. The ones you will use most:

• /init — Studies the project and writes LUCIAI.md at its root: instructions Lucicoder reads on every turn.
• /model <name> — Switches the model.
• /mode <mode> — Sets the approval mode: chat, ask, agent, auto or preview.
• /mcp — Lists your MCP servers and their tools. /mcp reload restarts them.
• /tasks — Lists the background tasks Lucicoder started, such as dev servers and watchers.
• /skills — Skills, commands and plugins: list, create and import them.
• /native auto|on|off — Native function calling, or the universal XML grammar for models that need it.
• /help — Shows every command.

Your own skills and Markdown commands appear in the same menu, as /their-name.

@file mentions

Mention a file with @ — for example @src/billing/total.ts — to put it in front of Lucicoder together with your message.

Sessions and context

Sessions are per project and per window. Each keeps its own agent memory, survives a restart of the IDE and resumes with its context.

Long sessions compress on their own, so they keep fitting the model's context window. Live language-server diagnostics are attached to every turn, so “fix the errors” just works.

For bigger jobs, Lucicoder keeps a live plan checklist, and can hand a focused part of the work to a subagent (spawn_agent) that has its own context.

Skills, commands and plugins

Lucicoder uses the layout Claude Code uses, so skill repositories, plugins and marketplaces made for it install unchanged — and what you write here works there.

Skills

A skill is a folder with a SKILL.md file — YAML front matter with a name and a description, then the instructions — and any files it needs. Lucicoder sees a one-line catalog of your skills and loads a skill's full text only when a task needs it; /name runs one directly.

Put project skills in .luciai/skills/<name>/SKILL.md inside the project, or in the same folder under your home directory for every project. Skills in .claude/skills are found too.

---
name: release-notes
description: Write release notes from the changes since the last tag
---
Read the Git history since the last tag. Group the changes into
Added, Changed and Fixed, in plain sentences, one line each.

Commands

A command is a Markdown file in .luciai/commands/ (or .claude/commands/). Its name becomes a slash command, and $ARGUMENTS is replaced by what you type after it.

---
description: Find and fix a bug from its description
argument-hint: <what goes wrong>
---
Find the cause of this bug, fix it and add a test that fails without the fix:
$ARGUMENTS

Plugins

A plugin bundles skills, commands and MCP servers in Claude Code's plugin layout. Manage skills, commands and plugins with /skills.

MCP servers

Model Context Protocol servers give Lucicoder more tools. Declare them in <project>/.luciai/mcp.json, in the format Claude Code and Claude Desktop use — you can copy an existing configuration as it is.

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "D:/some/dir"],
      "env": {}
    },
    "remote-example": {
      "url": "https://mcp.example.com/mcp",
      "headers": { "Authorization": "Bearer <token>" }
    }
  }
}

• A local server has a command, with optional args and env; it runs on your machine and talks over stdio.
• A remote server has a url instead, with optional headers, and is reached over HTTP.
• "disabled": true keeps a server in the file without starting it.

/mcp lists the servers and their tools, and /mcp reload restarts them after you edit the file. Lucicoder sees each tool as mcp__<server>__<tool>, and calls to them follow the same approval rules as commands.