Open Source Releases

spotwarp 3.2.2 — Keeps your spot GPU training runs alive by automatically backing up checkpoints and failing over to another cloud (Vast.ai, RunPod) in under a minute when instances get reclaimed. If you’ve ever lost a 3-day training run to a spot eviction, this is the insurance policy you wanted.

inspeximus 2.3.0 — A zero-dependency memory layer for agents that actually handles corrections: serve updated facts while suppressing stale ones, with value-ranked recall, per-type decay, and tamper-evident erasure. Includes an MCP server so your agents can plug it in directly. Finally, memory that doesn’t hallucinate its own history.

Coin Railz — Bundles 63 micropayment-gated services (crypto analytics, satellite data, IoT feeds, AI inference, prediction markets) behind a USDC x402 layer. Coinbase AgentKit compatible; pay per call or prepaid credits. The “everything store” for agent tooling, assuming you’re fine with on-chain micropayments for every API hit.

minirag-mcp 0.3.1 — Local-first RAG MCP server with hybrid vector + keyword search. No external API calls, no data leaving your box. Good for privacy-sensitive pipelines or when you just don’t want another managed service bill.

burmesenlp 1.0.1 — Production-grade NLP toolkit for Burmese: syllable/word/sentence segmentation, rule-based POS tagging, phrase chunking. Low-resource languages get ignored until someone ships real tooling — this is that.

AI Dev Tools

Understand-Anything — Turns codebases into explorable, searchable knowledge graphs that work with Claude Code, Codex, Cursor, Copilot, Gemini CLI. Focuses on “graphs that teach” over pretty pictures. Finally, a code graph that answers “where does this function get called?” instead of just looking cool.

TradingAgents — Multi-agent LLM framework for financial trading with specialized research, analysis, risk, and execution agents. A concrete example of coordinating agents for complex decision-making — useful reference even if you’re not building a trading bot.

Today’s Synthesis

Wire inspeximus’s correction-aware memory layer with minirag-mcp’s local hybrid search and Understand-Anything’s codebase knowledge graph, and you’ve got a fully local coding agent that remembers what it learned about your repo — and lets you fix its memory when it hallucinates. Point Understand-Anything at your monorepo to generate the graph, feed symbols and docs into minirag-mcp for retrieval, then wrap inspeximus as the agent’s long-term store: every “fact” (function signatures, architectural decisions, bug fixes) gets versioned, decay-ranked, and suppressible when you correct it. The MCP servers chain together — Understand-Anything exposes the graph, minirag-mcp handles retrieval, inspeximus handles persistence — so your agent (Claude Code, Cursor, whatever) queries one interface. No data leaves your box, no managed service bills, and when the agent confidently claims process_payment() lives in billing/v2 but you moved it last sprint, you delete the stale fact and the suppression propagates. That’s a coding assistant that actually improves over time instead of repeating the same mistakes. 🛠️