███╗   ███╗ ███████╗ ██╗      ██████╗
 ████╗ ████║ ██╔════╝ ██║      ██╔══██╗
 ██╔████╔██║ █████╗   ██║      ██║  ██║
 ██║╚██╔╝██║ ██╔══╝   ██║      ██║  ██║
 ██║ ╚═╝ ██║ ███████╗ ███████╗ ██████╔╝
 ╚═╝     ╚═╝ ╚══════╝ ╚══════╝ ╚═════╝

Agent-agnostic settings generator for AI coding agents. Define your projects, context, MCP servers, and skills once — generate native configs for Claude Code, Codex CLI, and Gemini CLI.

Install

Run directly without installing:

npx @vitld/meld-cli init
# or
pnpm dlx @vitld/meld-cli init

Or install globally:

npm install -g @vitld/meld-cli
# or
pnpm add -g @vitld/meld-cli

The global install exposes the meld command.

Quick start

mkdir my-hub && cd my-hub
meld init
meld project add
meld gen
meld claude-code   # or: meld codex-cli, meld gemini-cli

How it works

Meld creates a hub — a shared workspace that sits above your projects. You configure everything in one place (meld.jsonc), and meld gen produces native config files for each agent:

Agent	Generated files
Claude Code	CLAUDE.md, .mcp.json, .claude/settings.json
Codex CLI	AGENTS.md, .codex/config.toml
Gemini CLI	GEMINI.md, .gemini/settings.json

Meld also generates a .code-workspace file and manages .gitignore entries.

Hub structure

my-hub/
  meld.jsonc          # Central configuration
  context/            # Markdown instructions for agents
  skills/             # Reusable agent skills (SKILL.md per skill)
  artifacts/          # Research, plans, and notes
  scratch/            # Temporary work (gitignored)
  agents/             # Generated output (gitignored)

Configuration

All configuration lives in meld.jsonc at the hub root:

{
  "$schema": "./meld.schema.json",
  "ide": {
    "default": "cursor",         // "cursor" | "code" | "windsurf"
    "workspaceName": "my-hub"
  },
  "agents": {
    "claude-code": { "enabled": true },
    "codex-cli": { "enabled": false },
    "gemini-cli": { "enabled": false }
  },
  "projects": {
    "my-app": {
      "path": "/absolute/path/to/my-app",
      "aliases": ["app"],
      "repo": "org/my-app"       // optional — used in project table
    }
  },
  "mcp": {},
  "context": "./context/",        // optional — custom context directory
  "enable-external-skills": false  // optional — discover skills from .agents/skills/
}

Agents

Each agent supports these options:

Option	Required	Description
enabled	Yes	Enable config generation
dir	No	Custom subdirectory under agents/ (defaults: claude-code, codex, gemini)
overrides	No	Deep-merged into the agent's generated settings file

Overrides let you customize generated configs without editing output files. What you can override depends on the agent's native format:

"agents": {
  "claude-code": {
    "enabled": true,
    "overrides": {
      "env": { "CLAUDE_CODE_MAX_TURNS": "50" }  // → .claude/settings.json
    }
  },
  "codex-cli": {
    "enabled": true,
    "overrides": {
      "approval_policy": "never"                 // → .codex/config.toml
    }
  }
}

MCP servers

Defined once under mcp, automatically translated into each agent's native format (.mcp.json, .codex/config.toml, .gemini/settings.json).

Stdio server (local process):

"my-server": {
  "command": "npx",
  "args": ["-y", "my-mcp-server@latest"],
  "env": { "API_KEY": "sk-..." }
}

HTTP server (remote):

"my-server": {
  "type": "http",
  "url": "https://mcp.example.com/mcp",
  "headers": { "Authorization": "Bearer tok-..." }
}

Scoping to specific agents — by default all servers go to all agents:

"my-server": {
  "command": "node",
  "args": ["server.js"],
  "agents": ["claude-code"]
}

Context

Files in the root of context/ are inlined into agent instruction files (CLAUDE.md, AGENTS.md, GEMINI.md). Subfolders are copied into each agent's working directory so you can reference them with relative paths.

context/
  01-role.md          # Inlined (alphabetical order)
  02-guardrails.md    # Inlined
  reference/          # Copied as agents/<name>/reference/
    api.md
    patterns.md

Use numeric prefixes to control ordering. Set "context": "./my-context/" to use a custom directory.

Skills

Skills are reusable agent instructions with frontmatter metadata. Each skill lives in its own directory under skills/ with a SKILL.md file:

skills/
  deep-review/
    SKILL.md

Skills support YAML frontmatter for metadata like name, description, and per-agent model overrides. They are distributed to each agent's native skill format during meld gen.

Set "enable-external-skills": true to enable skills.sh support — this discovers third-party skills installed in .agents/skills/ and distributes them alongside your local skills.

CLI reference

Command	Description
meld init	Initialize a new hub
meld gen	Generate agent configs
meld gen --dry-run	Preview without writing
meld project add	Register a project
meld project list	List registered projects
meld open	Open workspace in IDE
meld update	Re-scaffold hub structure and regenerate
meld claude-code	Launch Claude Code in its agent directory
meld codex-cli	Launch Codex CLI in its agent directory
meld gemini-cli	Launch Gemini CLI in its agent directory

Team usage

meld.jsonc contains absolute project paths that differ per machine, so gitignore it — similar to .env. Everything else can be shared.

# ── meld managed (do not edit) ──
agents/
scratch/
# ── end meld managed ──

# Team additions
meld.jsonc

Commit	Don't commit
context/, skills/, artifacts/	meld.jsonc (machine-specific paths)
meld.schema.json (IDE autocompletion)	agents/ (generated, gitignored)
	scratch/ (temporary, gitignored)

Tip: Commit a meld.example.jsonc with placeholder paths as a template for new team members.

Contributing

Contributions are welcome! Try to keep it within the spirit of the project — a lightweight config generator, not an agent runner or orchestrator.

Known limitations:

• Windows is untested

Requirements

Node.js >= 20

License

GPL-3.0-only