[Feature request] Scoped, metadata-only API for AI agents managing collections/groups

[suantea]

Motivation

I'd like to let an AI agent help manage my vault organization (grouping collections, cleaning up structures, re-categorizing items). Today the only way to do this programmatically is the admin API, which is all-or-nothing: an agent with access to it can read/export everything.

Security concern

A vault server holds ciphertext (safe without the master password) but also plaintext metadata (item names, URLs, usernames, org/collection/group structure) that is itself sensitive. An agent integrating with a third-party LLM would ship this metadata off-box, and a broad admin token also turns prompt injection into a destructive capability (delete groups, change permissions, export).

Proposal

Add a scoped agent token / API surface with:

1. Fine-grained scopes (e.g. groups:read, groups:write, collections:read, collections:write) — never including ciphertext read or export.
2. Metadata-only endpoints that never return ciphertext fields, so agents only ever see names/IDs/structure.
3. Optional field redaction (names/URLs) for deployments that integrate external LLM APIs.
4. Audit events for every agent action, so operations are traceable and reversible.
5. Docs on running the agent with a local model (Ollama etc.) so data never leaves the host.

Alternatives considered

• Running a local model + a filtering proxy in front of the admin API — works today, but the token is still all-powerful and it's custom maintenance.
• Doing the reorganization client-side with a modified web vault — loses server-side automation and audit.

Questions

• Is a scoped agent API something the project would consider?
• Would a metadata-only read endpoint for collections/groups be acceptable as a first step?

---

中文说明（翻译摘要）

我想接入一个 AI agent 帮我管理保险库的分组/集合结构并做整理。目前只有 admin API 是"全有或全无"，agent 拿到就能读/导出一切，风险太大。希望提供：细粒度 scope（分组/集合的读写，绝不包含密文读取和导出）、只返回元数据（名称/ID/结构，不含密文）的接口、敏感字段脱敏选项、操作审计日志，以及本地模型部署文档。请问上游是否考虑支持这类受限 agent 接口？

[stefan0xC]

A vault server holds ciphertext (safe without the master password) but also plaintext metadata (item names, URLs, usernames, org/collection/group structure) that is itself sensitive.

If you look at the entries in your database you will see that much of the metadata is not saved in plaintext but is stored encrypted as well. Given that the clients are responsible for decrypting the data this feature request is probably out of scope for this project as it adds too much of an overhead.