The interesting part is not the cache but the expiry policy: a Next.js answer is treated as stale far sooner than a TCP answer, so reuse does not quietly turn into wrong. Read the privacy section before you turn it on — what leaves your machine is a sanitised technical summary, and the server hard-rejects anything containing local file paths, but the contributions are public by design.
A shared research cache for coding agents. Before the agent hits the web, wellread looks for an existing verified summary; on a miss it lets normal research happen and stores the result for whoever asks next. It works on its own from day one — your own past research comes back to you — and improves as more people contribute.
- A hit returns an answer built from verified sources, with no web searches at all
- A partial hit starts from what exists and researches only the gaps
- A miss falls through to normal research, then saves the summary
- Freshness is per topic: timeless entries stay fresh for a year, stable ones for 6 months, evolving ones for 30 days, volatile ones for 7 days — and a re-verification resets the clock for everyone
- Ask for your wellread stats to see token savings and how often your saved research was reused
- Six privacy layers, including server-side hard rejects for any content containing local paths or non-http sources
The npm package `wellread`, version 0.1.36. Run `npx wellread` and restart your editor; `npx wellread@latest` updates and `npx wellread uninstall` removes it. It works with any MCP client, with Claude Code called out as the best-supported, alongside Cursor, Windsurf, Gemini CLI, VS Code and OpenCode. Licensed AGPL-3.0.
One command — npx wellread