agentFramework note 1
Filter: orchestration stays in n8n, routing in LiteLLM, state in Supabase, building in Claude Code. A framework must do something that stack cannot.
agentFramework
Ordered by verdict, then GitHub stars.
Filter: orchestration stays in n8n, routing in LiteLLM, state in Supabase, building in Claude Code. A framework must do something that stack cannot.
agentFramework
dspy is the standout: the golden set + accuracy guard give it a pass/fail metric to optimize local Qwen prompts against.
agentFramework
e2b gives generated code a disposable sandbox instead of running beside Sierra Chart or next to keys on the laptop.
agentFramework
Speech-to-text is a genuine gap: local Whisper on the RTX 4060 first, Deepgram as the paid fallback; feeds ESP32-S3 recorder audio and consulting call transcripts.
agentFramework
agentops adds run traces that model_calls lacks.
agentFramework
Risk: crewai, autogen and langchain each want to own orchestration; running one alongside n8n means two drifting descriptions of the same workflow.
agentFramework
The post's '776K combined stars' claim is not supported: humanizer 31.5k and napkin 575 verified on GitHub 2026-09-17; the other 18 links were truncated and are unverified.
agentSkill
Skills add procedure; MCP servers add tools. Our stack is tool-heavy and procedure-light.
agentSkill
Install safety: read SKILL.md fully, install per-project, pin a commit, never load unvetted skills in a session holding Supabase service keys or API keys, review diffs before updating.
agentSkill
Humanizer improves style but does not remove KDP/Etsy duties to disclose AI-generated content.
agentSkill
Adoption order: napkin, humanizer -> ffmpeg-skill, video-shotcraft -> graphify, tech-debt-audit -> claude-seo, scroll-world, user-research.
agentSkill
Verified 2026-09-17: obra/superpowers ships via Anthropic's official plugin marketplace; steveyegge/beads is a git-backed issue tracker for agents. The other 38 were not individually verified.
agentToolkit
Adopt order: superpowers, beads, serena, repomix, planning-with-files, karpathy-skills, codex-plugin-cc, taste-skill, addyosmani skills, mattpocock skills, claude-hud, rtk.
agentToolkit
Risk: headroom/rtk/caveman compress context or output; the accuracy guard rejects numbers missing from source, so never use them on the daily data pipeline or publishing paths.
agentToolkit
Risk: terse-output tools strip disclaimers required for finance, health-adjacent and gambling content.
agentToolkit
Connect the MCP and ask Claude to search this collection for you.