Questions teams ask us.

The Model Context Protocol is the open standard AI tools use to reach external knowledge. Because Verde speaks standard MCP over HTTP, it works with Claude, Claude Code, Cursor, Codex, ChatGPT, and any future MCP-compatible client, so you're never locked into one vendor's plugin system.

Yes, if you grant it that. When the connected token belongs to a member with publish permission, the AI can take Public documents live. That is deliberate, since some teams want their agent keeping a public KB current. Two things are enforced on our side: a public document never goes live as a side effect of being written (that always leaves a draft), and the publish call is refused unless it carries an explicit confirmation flag. What we cannot enforce is whether the agent really showed you the content before setting that flag. The tool description instructs it to, but that's the agent following instructions, not a guarantee. So treat publish permission like any capability you hand an agent: grant it only for clients you trust. Without it, public writes are saved as drafts for a human publisher, and no AI client can widen a document's visibility.

No. Private memories belong to one member and are invisible to everyone else, including owners and admins. The rule is enforced in SQL on every query and verified by our end-to-end test suite.

No. Retrieval uses weighted Postgres full-text search, so there's no AI bill attached to search. The search service is isolated, so embeddings can be added later without a rewrite.

You supersede the old memory with the new one. History is preserved and clearly marked, but normal retrieval, whether human or AI, only returns the current version. Tools can explicitly ask for historical context when they need it.

Each member generates their own bearer token (shown once, stored as a SHA-256 hash). A token acts with exactly that member's permissions and can be revoked at any time. OAuth 2.1 is supported for clients like Claude.ai that prefer it.