Metadata Auto Classifier

Automatically classifies and applies metadata to your notes.

Features

• Selects existing values for tags and other frontmatter fields with calibrated yes/no decisions: each candidate gets its own probability, and only confident yeses are written.
• Classification uses decision API models only: TypeSafe Jev, OpenAI Decisions, OpenRouter or a custom /systemone endpoint. Chat models (Anthropic, Gemini, OpenAI, Codex, Ollama, …) never classify; they only generate new tags.
• Generates new values only for the exact tags field when Generate new tags is explicitly enabled; this is off by default.
• Preserves field instructions, references, minimum/maximum counts, overwrite/append and WikiLink options. Writes preserve unrelated frontmatter and note text.
• Providers include TypeSafe, OpenAI, Anthropic, Gemini, OpenRouter, Ollama, DeepSeek, LM Studio, Codex (OAuth), and an explicitly configured OpenAI-compatible custom endpoint.

Usage

1. Open Settings, Metadata Auto Classifier, add a provider and a decision API model (mark it in the model editor; TypeSafe models always are). Add a chat model too if you want new tags.
2. Define the fields, existing reference values, instructions and count limits. For optional new tags, edit the exact tags field, enable Generate new tags, New tags use the selected model when it is a chat model, otherwise the first chat model. Local Ollama and LM Studio can be keyless.
3. Open a note and run Fetch frontmatter: Fetch all frontmatter using current provider from the command palette, or Fetch frontmatter: tags (or another field) for one field.
4. Review the result in the note's properties.

Ordinary fields never generate new values. Ambiguous selection or a failed required generation request leaves the field unchanged. The maximum count limits values selected/generated in that operation, not the total values already present when appending. Empty candidates skip the classification request; no remaining generation slots skip generation.

Installation

Community plugins: open Settings, Community plugins, Browse, search for "Metadata Auto Classifier", then Install and Enable.

Manual: download main.js, manifest.json and styles.css from the latest release into <vault>/<configDir>/plugins/metadata-auto-classifier/, then enable the plugin. <configDir> is your vault's configuration folder (.obsidian by default).

Disclosures

• Network use: selection sends the note body, instructions and candidate values to the selected provider: api.typesafe.ai/v1/systemone for TypeSafe, <base URL>/decisions for OpenAI decision models, <base URL>/systemone for other decision models. Explicitly enabled tag generation sends the note body, instructions and excluded/existing values to the chat generation model. Reference values may come from metadata on other notes; those other note bodies are not sent by classification.
• Local usage records: token counts and response times from calls you already make are stored per provider in <configDir>/plugins/metadata-auto-classifier/usage.json and shown under Providers. They cover this plugin's requests only, not everything a key is used for. Note text is never stored, and no extra API calls are made to collect them.
• Account or payment: each provider uses its own credential. Settings from 1.13 that stored a separate Jev key become a selected TypeSafe provider. Providers bill usage according to the selected account/model; model presets do not guarantee availability or pricing. API keys and OAuth tokens are stored in the plugin's settings file; do not commit or share that file.
• Codex sign-in: hosted device authorization contacts auth.openai.com; Codex inference contacts chatgpt.com. The verification page opens only after an explicit click. There is no local callback server or automatic clipboard copy. Refresh remains account-bound and must be saved before retrying.
• Transport: requests use Obsidian's native requestUrl, not Node HTTP streaming. Cancellation prevents stale processing/writes but cannot physically cancel a request already dispatched by that API.
• Folder contents: the note picker for reference values lists one folder at a time as you browse. When a field has no reference values, the plugin reads the tags or property values used by notes in the classified note's own folder. It never scans the whole vault, and nothing beyond the note being classified is sent to a provider.
• Telemetry: none.

Offline tests do not certify live-account entitlement, mobile execution or semantic classification quality.

Development

See the dev contract: bun run dev, bun run build, bun run ci.

Install with bun install --frozen-lockfile. Builds go to dist/; workspace tooling owns development-vault deployment. Shared CI/release workflows remain referenced from the upstream template.

License

MIT