Hi Product Hunt 👋 I'm Ismail, I built briefd on weekends while running many projects in one domain at work. The problem: teams keep their rules, glossary, architecture decisions and conventions in CLAUDE.md / AGENTS.md files, and those get loaded whole into every agent session on every turn. You either pay for the whole library each turn or keep the file small and leave most of the knowledge out - and every project copies the same domain knowledge. briefd flips it: knowledge lives once, as Markdown in a git repo, and the agent asks per task - "what do I need to know for this?" - over MCP. It gets a compiled bundle that never exceeds the token budget it set. Agents can't edit the knowledge; they propose changes as pull requests and a human merges. What makes it more than retrieval: every section carries the date its document last changed; point it at your code repos and it tells the agent when the code a rule governs changed after the rule was written; and the questions agents asked that the corpus couldn't answer show up on the dashboard as a backlog. Measured on the sample corpus: −86% knowledge tokens per task, right section present 96% of the time; in real Claude Code sessions -40% cost with identical answers. Benchmarks and a 90-second explainer are in the README. Tech: one Go binary, one SQLite file, hybrid BM25 + embeddings - the embeddings run in a from-scratch pure-Go BERT encoder (no ONNX runtime, no CGO), multilingual by default. Works with Claude Code, Cursor, Codex and any MCP client, over HTTP (team) or stdio (solo). briefd demo runs the sample knowledge base in 30 seconds. I'd love to hear how you manage shared knowledge for agents today, and what would make you switch. Apache-2.0: https://github.com/ismailperim/b...

briefd

Token-budgeted team knowledge for coding agents, from git

Token-budgeted team knowledge for coding agents, from git