Hey PH! Solo dev here. Shinobi started because I kept re-explaining the same decisions to Claude Code every session. Every conversation began with "remember we picked Postgres, not MongoDB?". The model is brilliant in a single session and a goldfish across sessions. This is the MCP server I built to fix it. Local-first SQLite, plug into any MCP client (Claude Code, Cursor, Cline, Continue.dev, Zed), your data never leaves your machine unless you opt into the hosted SaaS (coming later). Install: npm install -g @shinobiapps/shinobi What's inside: • Projects + tasks the agent claims/completes across sessions • Decisions log with rationale (memory across sessions) • Dead ends — failed approaches the agent doesn't retry • Notes, voice notes, versioned plans, per-project context • Web dashboard at localhost:8765 with Cmd+K command palette • Cross-machine git-based sync, no cloud account required Happy to answer any questions about the MCP design, the Cloudflare Worker relay for multi-agent sync, or the embedding-provider abstraction (Ollama/Voyage/OpenAI).

Shinobi

Local-first memory layer for AI coding agents

Local-first memory layer for AI coding agents