Hello,
Sharing a small open repo: GitHub - sbknext/forge-frappe-skill · GitHub
Why I built it:
AI coding agents work well with good context. The strongest ones (e.g. Claude Code, Codex)
handle Frappe reasonably, but most other / cheaper IDE + agent setups lose the framework
context on custom apps and generate outdated patterns — pre-v13 code, raw SQL instead of
Query Builder, missing permission checks, wrong hook usage. I wanted that Frappe knowledge
as something any agent can search + load on demand: DocTypes, controllers, hooks, Query
Builder, permissions, REST/RPC APIs, reports, testing, bench/ops.
The point is access: with these skills + free or self-hosted models (Kimi / Qwen / DeepSeek
via OpenRouter, Groq, or local), a student or a small business can build real Frappe apps
without depending on one expensive agent — near-zero API cost, and the same Frappe context
the best tools get. The direction around it: a searchable code-graph + secure-scan + test
pipeline for custom apps.
What’s in it right now:
- ~471 granular reference skills + 11 installable core skills (plain SKILL.md / markdown)
- Works via any MCP host or fully standalone (just the .md files — no lock-in)
- Frappe-accurate (v13–v15): Query Builder, whitelist + permission checks, hook-based controllers
Credit — it builds on existing OSS, thanks to:
- Impertio-Studio/frappe-skills, lubusIN/frappe-skills, the official Frappe docs, and netchampfaris, and google web scrap + claude code opus 4.8 verification
Sharing it so people can actually use it — and so it improves on its own as the community
adds and corrects skills. PRs, corrections, and missing topics welcome.
Note / tip: This repo is maintained with an AI harness in the backend, so you’ll
see frequent auto-improvements (skills getting added + corrected over time). Because of
that, treat it as evolving — not recommended for production code as-is.
Good path:
build and iterate on demo apps at any scale first; once you’re confident, point it at a
testing system to automate your tests; and eventually fold the verified pieces into prod
via server scripts, client scripts, and other customizations. Use judgement, review the
output — it’s context for your agent, not a substitute for review.
Thanks,
Sambhaji K