# HaxyScott Technologies

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Christian Lurhakwa](https://cerebralvalley.ai/u/irchriscott)
- **GitHub:** https://github.com/irchriscott/reincore
- **Demo video:** https://www.youtube.com/watch?v=Ynyy3Pkdc8A
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1301

ReinCore is a local gateway that makes AI agents cheaper, safer and harder to derail, without stopping them mid-task. Its rule is simple: guided, never stopped.

The problem

AI coding agents like Muse Code, Claude Code and Codex don’t make one LLM call. They make dozens or hundreds, all working toward one goal. Today’s gateways and spend controls treat each call as an isolated request, and that causes three problems:

- Wasted money. Agents resend the same context and reread the same files. They repeat the same tool calls and get stuck in retry loops. Small changes in the prompt also break the provider’s prompt cache, so its discounts are lost.
- Lost work. Budgets act as hard walls. When a spending cap or rate limit hits, the task dies mid-run and the progress is gone.
- No safety net or visibility. Tool calls run unchecked, and secrets and personal data flow to providers. Prompt injections hidden in tool output go unnoticed. Developers can’t easily see what a session cost or whether it actually succeeded.

The solution

ReinCore runs on your machine, between the agent and the LLM providers and tool servers it calls. Setup is one change: point the agent’s base URL at ReinCore. Because it treats the whole session as the unit of work, it can:

- Cut cost. It stabilizes prompts so provider caching works and reuses the results of read-only tool calls. It compacts long histories and routes each step to the cheapest model that can handle it, with routine steps going to Muse Spark 1.3 and hard reasoning to Muse Wonder 1.0.
- Get cheaper instead of stopping. As a budget runs low, sessions move to cheaper models. If a pause is truly needed, ReinCore saves the session and returns a token that resumes exactly where it stopped.
- Keep agents on track. It detects loops, malformed or unknown tool calls, prompt injection and unsupported citations. It redirects the agent instead of killing the session.
- Protect data. Every tool call passes an action gate. Secrets and personal data are masked before they leave the machine, and everything is stored locally in SQLite.
- Show what happened. Each session gets cost receipts, an exportable timeline, and separate records of whether the task completed and whether it succeeded.

ReinCore is written in Rust and ships as a CLI (reincore / rc), a local daemon, an MCP gateway and a desktop app with dashboards for projects, sessions, budgets and providers. It works with Muse Code, Claude Code, Codex, Cursor, Gemini CLI, Aider, GitHub Copilot and LangGraph/CrewAI/AutoGen agents. reincore launch muse runs Muse Code through ReinCore on Muse models, end to end.

## More from Meta Global AI Developer Hackathon

- [AutoLab](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1298)
- [Team Bliss](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1299)
- [Team Eco cycle](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1300)
- [Pixel Heist](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1302)
- [Reel Studio](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1303)
- [verifiable-archive-indexer](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1304)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1301. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt