SkillSinc

Write your agent skills, instructions, agents, prompts, MCP servers, reference files, and hooks once. SkillSinc copies them into Claude Code, GitHub Copilot, VS Code, Codex, and the shared ~/.agents folder, converting formats where it must and saying so when a conversion loses something.

It replaces hand-copying, symlinks, and per-tool sync scripts with one command that shows you every change before it makes it.

How it works

graph LR
    Library[Your library<br/>skills/ instructions/ agents/ prompts/] --> Plan[skillsinc plan]
    Plan --> Claude[~/.claude]
    Plan --> Copilot[~/.copilot]
    Plan --> VSCode[VS Code prompts]
    Plan --> Codex[~/.codex/AGENTS.md]
    Plan --> Agents[~/.agents/skills]

• Your library is the source of truth. A sync makes each destination match it: new items are added, changed items updated, removed items deleted.
• SkillSinc only touches what it wrote. It records every file it owns. An existing file it did not write blocks that destination until you review it and pass --adopt.
• Nothing is written without a plan. skillsinc plan lists every change; skillsinc sync applies the same plan.
• Lossy conversions are opt-in. If a harness cannot represent a field (for example, Claude ignores Copilot tools lists), the item is skipped until the route sets allow_lossy = true.
• A broken source never deletes anything. If a library cannot be read completely, its destinations are left untouched.

Install

Requires Python 3.12+ and uv.

uv tool install git+https://github.com/mikejhill/skillsinc.git

Quick start

1. Write a configuration. It enables a destination for every harness folder found on this machine and creates an empty library at ~/.config/skillsinc/library ($XDG_CONFIG_HOME/skillsinc/library). Pass --source DIR to use an existing library instead:

   skillsinc init
   

2. Preview the changes:

   skillsinc plan
   

3. Apply them:

   skillsinc sync
   

Run skillsinc sync again whenever the library changes, or schedule it with Task Scheduler, launchd, or a systemd timer using the scheduling examples. An overlapping run exits without writing.

Library layout

The layout matches GitHub Copilot's file conventions, so an existing Copilot library works unchanged:

my-agent-library/
├── skills/<name>/SKILL.md            # plus any supporting files
├── instructions/<name>.instructions.md
├── agents/<name>.agent.md
├── prompts/<name>.prompt.md
├── mcp/<name>.json                   # one MCP server definition per file
├── resources/<name>.md               # reference files, copied as written
├── hooks/<name>.json                 # one Copilot CLI hook file per file
└── plugins/<plugin>/                 # Claude Code plugin layout, expanded into the kinds above

INTENT.md, *.intent.md, and hidden files are never distributed. A plugin's skills/, agents/, commands/, and .mcp.json are synced as skills, agents, prompts, and MCP servers; everything else in it (hooks, scripts, LSP servers) is reported as a loss. See decision 0005.

What each harness receives

Run skillsinc profiles for the same table. See harness profiles for exact conversions.

Commands

plan and sync accept --destination NAME (repeatable), --adopt, --json, --verbose, and --config PATH. Exit codes: 0 success, 1 a destination was blocked or failed, 2 configuration or usage error, 130 interrupted.

Documentation

• Configuration reference
• Command-line reference
• Scheduling
• Development guide
• Roadmap
• Harness profiles
• Philosophy and goals
• Architecture
• Decision: rebuild scope
• Decision: MCP server entries
• Decision: resources and Claude imports
• Decision: TOML entries with tomlkit
• Decision: plugins expand into kinds
• Decision: entry files beside a destination root
• Decision: edit VS Code's JSONC mcp.json in place
• Decision: one copy of each item per harness
• Decision: MCP server options and per-profile overrides
• Contributing · Security · Changelog

License

MIT