sAIdecar (pronounced "sidecar") is a new terminal-based AI tool that fills a gap most developers eventually hit: the small questions that come up during coding work but don't belong in your main agent session or flagship LLM context. The project, built by developer Deca and shared on GitHub this week, is a Node.js CLI running on Node 22 or newer. It opens an interactive prompt for quick Q&A while you keep working, stores every exchange as append-only JSONL logs locally, and indexes everything into SQLite so you can search your question history later by keyword, tag, saved status, or date.

The Core Philosophy: Stay in Your Lane

The key differentiator is what sAIdecar *doesn't* do. It doesn't inspect repositories, edit files, run shell commands, or mutate your project in any way. That's the whole point—it's a question-and-answer companion with local-first memory, not another AI agent trying to take over your workflow. You keep it open in a narrow terminal pane beside your editor (Zellij is the recommended setup), ask quick clarifying questions as they come up, and move on without breaking focus or context. The conversations are written as append-only JSONL files under ~/.saidecar/, with a rebuildable SQLite FTS5 index for search. If you nuke the database, its from the logs—nothing is lost.

Multi-Provider Support

The tool supports multiple backends: OpenAI (direct API), Codex CLI auth (Sign in with ChatGPT), Anthropic (Claude), DeepSeek (OpenAI-compatible path), and MiniMax. Configuration is straightforward environment variable work—you pick one provider per session and go. For web-backed questions, sAIdecar can use either provider-native search or a self-hosted SearXNG instance. Public instances are available but rate-limited; Deca notes that running your own via Docker is more reliable and private for heavy usage.

Search, Tags, and Digest Reviews

The saidecar-logs command gives you a read-only terminal UI to browse your archive with full-text search, filters by project/date/backend/mode, tag-based browsing, favorites, and importance badges. Entries containing code snippets get special markers, and high-importance questions can be flagged for later review. For weekly reflection, saidecar-digest generates deterministic Markdown reviews from your logs—daily or weekly summaries of what you asked about and learned. Optional LLM-written narrative summaries are available if you want the tool to add context beyond just listing entries.

Safety by Design

With SAIDECAR_AUTO_FILTER enabled, a lightweight background model scores newly logged entries as keep/condense/discard based on usefulness—trivial questions get summarized rather than stored in full, while decisions, references, and code snippets are preserved. Manual saves via /save always keep the complete entry regardless of auto-filtering.

Key Takeaways

  • sAIdecar is a focused terminal scratchpad for quick AI Q&A—it does not touch your codebase or run autonomous tasks
  • Local-first storage with JSONL logs as source of truth and SQLite index that can be rebuilt from scratch
  • Supports OpenAI, Codex CLI, Anthropic, DeepSeek, and MiniMax providers
  • Built-in search, tagging, favorites, and daily/weekly Markdown digest generation
  • Available under MIT license at github.com/Deca/saidecar

The Bottom Line

This is exactly the kind of narrow-focus tool that hacker culture does best—solving your own itch with software that does one thing well. If you've been cluttering up your coding agent sessions with quick clarifications and syntax checks, sAIdecar gives you a cleaner separation of concerns without adding yet another SaaS subscription or cloud dependency. It's local, it's auditable, and it stays out of your way until you need it.