Offsend

About Offsend

Offsend is an open-source, local-first boundary layer that decides which files in a repository are allowed to enter AI context. The premise is that coding agents pull context through far more paths than the prompt box — they read workspace files, follow file references, run shell commands, call MCP tools, spawn subagents, and keep local transcripts — so a boundary has to cover all of those, and it has to cover uncommitted files too. The mechanism is a single `.offsend.yml` policy file that lives in the repo, so the rules are visible, reviewable in a PR, and versioned alongside the code rather than living in someone's editor settings. The CLI workflow is four commands: `offsend show` lists which sensitive paths may currently enter AI context, `offsend init` writes the policy, `offsend protect` adds the missing boundaries, and `offsend hook install` installs runtime gates. Those gates check prompts, file reads, shell commands, MCP arguments, and local transcripts for Cursor, Codex, Windsurf, and Claude Code, while ignore-file support covers GitHub Copilot, Continue, Gemini, LLM, Aider, Cline, Roo, Zed, and Cody. There is also `offsend check` for content scanning of keys and tokens in files, staged diffs, and stdin. The vendor is refreshingly precise about limits: it is defense-in-depth, not a guarantee, it is not a replacement for Gitleaks or TruffleHog on git history, and the CLI is explicitly labeled a research project not ready for production. Nothing is uploaded — checks run locally. Companion surfaces include a macOS app, a browser extension that masks sensitive values before browser prompts, a GitHub Action, and Radar, a research report on AI-context risk across public repos.

Key Features

Tags