Do not open a public issue for security vulnerabilities.

DM us on X (@codependent_ai) or message the Telegram channel with:

• Description of the vulnerability
• Steps to reproduce
• Impact assessment (what an attacker could do)

We'll acknowledge within 48 hours and aim to patch critical issues within 7 days.

resonant-archive is local-first software that runs on your machine:

• No cloud backend — the archive is a single ChromaDB directory at ~/.resonant-archive/ by default. Your data stays local.
• No telemetry — nothing phones home. The only network traffic is the one-time embedding model download from HuggingFace on first use.
• No authentication — the daemon binds to 127.0.0.1 only. It's not designed to be exposed over a network.
• No API keys required — embeddings run locally via sentence-transformers. There's no third-party service being called.
• Your content is your own — indexed text stays in your local Chroma store. We can't see it.

• Do not expose the daemon port (8766 by default) over the public internet. It has no authentication and isn't designed for remote access. If you need remote access, put a reverse proxy with authentication in front of it.
• Be careful what you index. If your notes contain secrets, API keys, passwords, or private third-party data, those will land in the Chroma store and be retrievable via semantic search. resonant-archive preserves everything it's given — it's not an access-control layer.
• The daemon writes to disk. The default location is ~/.resonant-archive/. On shared machines, make sure the directory has appropriate permissions.
• Do not commit ~/.resonant-archive/ to git. The .gitignore in this repo catches common copies of it, but be aware it exists in your home directory by default.
• Model cache location — sentence-transformers downloads models to ~/.cache/huggingface/ (or platform equivalent). Treat the cache like any other ML model cache.

• Leaking your source files. If your Obsidian vault has secrets and you index it, those secrets are now in the archive. Curate what you index.
• MCP clients with broad permissions. If you expose archive_search to an AI that can then be prompt-injected, a hostile prompt could exfiltrate archive content via tool-call output. This is a general MCP concern, not specific to resonant-archive.
• Filesystem compromise. If an attacker has shell access to your machine, they can read the Chroma database directly and don't need resonant-archive at all.

We accept reports for issues in:

• The resonant_archive Python package (chunker, embedder, store, daemon, MCP server, CLI)
• Documentation that misleads users about security behavior
• The default configurations (listen addresses, permissions, paths)

We don't control and can't fix:

• Bugs in chromadb, sentence-transformers, torch, fastapi, or other upstream dependencies (report those to their respective projects, but feel free to mention them to us so we can track)
• Issues in Claude Desktop, Claude Code, or other MCP clients
• Issues in Nexus AI Chat Importer (report upstream)