Skip to content

README.md

Latest commit

 

History

History
59 lines (50 loc) · 2.71 KB

README.md

File metadata and controls

59 lines (50 loc) · 2.71 KB

agent-toolkit

Portable, single-source-of-truth repo for my DIY agent skills + MCP servers, with a one-command installer that sets them up across Claude Code, Codex, Cursor, and OpenCode — each in its own native config, idempotently, never overwriting tokens and never storing secrets.

Inspired by anthropics/skills, extended to also carry MCP server definitions and a multi-agent installer.

What's inside

skills/
  mineru-sciverse/    # high-fidelity PDF→Markdown (MinerU) + literature search (SciVerse)
  learning-mechanics/ # applied "learning mechanics" DL-engineering reference (from arXiv 2604.21691)
mcp/
  servers.json        # canonical, agent-neutral MCP server defs (source of truth)
install/
  install.sh          # detect agents → copy skills → apply MCP (native format) → fill-token notes
  apply_mcp.py        # translates servers.json into each agent's config (idempotent)
docs/
  COMPATIBILITY.md    # per-agent skill/MCP paths + which agents share vs. stay separate

Quick start (new device)

git clone <this-repo> ~/agent-toolkit
cd ~/agent-toolkit
bash install/install.sh                 # auto-detects installed agents
# optional: also install local MinerU CLI (~2GB, offline parsing)
bash install/install.sh --with-mineru-cli

Then fill the two tokens in each agent's own config and restart that agent:

Useful flags: --agents claude,codex · --dry-run.

All four agents implement the Agent Skills standard and cross-read each other's home skill dirs, so the installer only populates ~/.claude/skills (read by Claude/Cursor/OpenCode) and ~/.agents/skills (read by Codex/Cursor/OpenCode) — never per-agent duplicates. MCP config is the opposite: each agent has its own file/format, so it's written four separate ways. See docs/COMPATIBILITY.md.

Design rules

  • One source of truth (mcp/servers.json); the installer fans it out per agent.
  • Each agent in its own lane — only ever writes its own config file. No cross-writes (this is what prevents one agent from clobbering another's setup).
  • Idempotent — re-runnable; existing servers and tokens are left untouched.
  • No secrets in git — tokens are written blank and filled per device; .gitignore blocks real-config filenames as a backstop.

See docs/COMPATIBILITY.md for exact paths and the share-vs-separate matrix.

Verify

bash skills/mineru-sciverse/scripts/check.sh   # status of local tools + tokens
python3 install/apply_mcp.py --agent all --home /tmp/sandbox   # safe dry test