// the find
ruvnet/agentic-flow
Easily switch between alternative low-cost AI models in Claude Code/Agent SDK. For those comfortable using Claude agents and commands, it lets you take what you've created and deploy fully hosted agents for real business purposes. Use Claude Code to get the agent working, then deploy it in your favorite cloud.
A scaffolding tool that sets up your Claude Code project with pre-written agent prompts, slash commands, hooks, and an MCP server via `npx agentic-flow init`. The actual core value is the .claude/ directory it generates — 80+ agent markdown files, 100+ command definitions, and hooks that fire before/after tool use. The rest of the README is marketing theater wrapped around what is fundamentally a prompt library.
- The hook system is genuinely useful: pre-edit and post-edit hooks that record outcomes to SQLite and surface agent suggestions are a practical extension of Claude Code's hook API that most users won't build themselves.
- 213 MCP tools wired up out of the box means you don't have to write boilerplate MCP server code for common GitHub, swarm, and memory operations.
- The `pretrain` command that walks git history and file structure to bootstrap routing patterns is a smart cold-start solution — most similar tools just give you static templates.
- TypeScript throughout with a CLI-first design means it's easy to inspect what it's actually doing, and the SQLite persistence layer for learned patterns is simple and auditable.
- The benchmark numbers are fabricated or meaningless. 'Flash Attention' reporting 0.00ms latency is measuring JavaScript array math, not GPU attention kernels. The '84.8% SWE-Bench solve rate' and 'Grade A+ Perfect Integration' are self-awarded. None of this is independently verified.
- The 'self-learning' and 'LoRA fine-tuning' claims are SQLite text storage with cosine similarity — there is no actual model training happening. Calling this 'SONA: Self-Optimizing Neural Architecture' is misleading to anyone who knows what LoRA actually is.
- The repo is a content farm: hundreds of near-identical markdown files with slightly different names (coder.md, dev-backend-api.md, implementer-sparc-coder.md) that are all variations on the same agent prompt template. Maintaining or customizing this at scale would be painful.
- Hooks that run `npx agentic-flow@alpha` on every file edit add a process-spawn overhead to every tool call in Claude Code. For a large session with many edits, this adds up, and the alpha tag on a dependency that runs before every write is a reliability risk.