The software
Free and open source, with no plan, no usage meter, no per-item billing and no paid team tier to unlock later; the project states that as a commitment rather than a launch price.
One Worker, a D1 database (SQLite), a Vectorize index, Workers AI for the embedding model and generation, and a KV namespace run it, all in your own Cloudflare account. The project's maintainers cannot read your memories or credentials, and Second Brain has no way to bill you, because nothing of yours ever passes through anyone else's hands.
Cloudflare usage
The project's stated position is that Second Brain runs within Cloudflare's free tier at personal scale. For teams, the wiki adds that a small team can often stay within free-tier budgets, and suggests monitoring your Cloudflare usage as the number of people and memories grows. There is no enforced limit on team size in the product; the practical limit is your plan's budgets, and a handful of people is comfortable.
The only way a charge could occur is usage in your Cloudflare account exceeding what its plan includes for free, at which point Cloudflare's own pricing applies. Captures and recall both draw on Worker requests, D1 queries, vector operations and Workers AI inference, so if a bill ever arrives, it comes from volume: many people, many memories, heavy recall. It would come from Cloudflare, never from Second Brain.
Neither the wiki nor the README states specific free-tier figures, such as request counts, database rows or vector counts. Cloudflare sets and changes those numbers itself, so check Cloudflare's current Workers, D1, Vectorize and Workers AI pricing rather than trusting any figure quoted elsewhere.
AI-client context
There is a second kind of cost that is paid in tokens rather than money. A recall at the start of a conversation adds roughly 500 to 2,000 tokens depending on how many memories come back, which is small against a 200k context window, and topK can be tuned down. Store calls are effectively free. The habit worth avoiding is letting an assistant store its own response after every turn; storing only the reusable ones, decisions and solutions, keeps things lean. The wiki's own verdict for most conversations: neutral to slightly positive, because one targeted recall replaces re-pasting context by hand.