Settings & local models

Open Settings from the top navigation. It's split into sections; this page covers the ones you'll use most. (Settings writes your vault's .anthracite/config.json for you — see Core concepts.)

Settings sections at a glance

Section	What it's for
Vault	Create/switch/add vaults (prompts a guided backend restart on switch); open `ANTHRACITE.md`; re-run onboarding
Connections	reMarkable, Readwise, Notion, Apple Notes, Folder watch, Obsidian (ingest-only)
MCP Servers	Connect external MCP servers to pull in documents
API Keys	Provider keys (OpenAI, Anthropic, Google Vision, …) — stored in the system keychain
Ingestion & OCR	LLM + OCR provider/model selection, languages, automatic-processing behavior, and the Ollama one-click install
Voice	Local transcription engine/model (Whisper variants)
Developer	Local ingest token + webhook URL/secret + bound port

Local models — free, private, one click

Anthracite works with bring-your-own API keys, but you can run everything locally and for free — no key, no upload. Two local backends are supported:

Ollama (one-click install)

Open Settings → Ingestion & OCR and find the local-model card:

Ollama running: the card shows a green status with the version and how many models are installed.
If the recommended model is already installed, you're done.
If not, click Install <model> — Anthracite pulls it for you and shows a live progress bar.
Ollama installed but not running: start the Ollama app, then click Re-check.
Ollama not installed: click Download Ollama to get it, then re-check.

The same one-click install also appears in the onboarding wizard's Local model step, so you can be running on a free local model before you finish setup.

MLX (Apple Silicon, fast OCR)

On Apple Silicon, MLX-VLM with GLM-OCR (or DeepSeek-OCR) is the recommended path for reading and processing. If you're running from source, see the root README.md for scripts/download-mlx-models.sh and the mlx_vlm.server command; then select the MLX provider/model in Settings → Ingestion & OCR.

Bring your own key

Prefer a hosted model? Add the provider key under Settings → API Keys (stored in the macOS keychain), then pick the provider/model in Ingestion & OCR. Keys never go into config.json.

Ingestion & OCR — the knobs that matter

Beyond model selection, this section holds the behavior settings you'll actually want to tune:

Setting	Default	What it does
Default OCR language	`de+en`	Language hint for reading new imports and reMarkable pages (changeable per page in Review via the Handwriting language selector)
Processing output language	match source	Language the wiki is written in
Answer language	match question	Language Ask replies in
Process automatically (read and organize new files after sync)	off	Read + process sources as they arrive (see After capture for which connectors participate)
Process stale sources overnight (local LLM, quietly at night)	off	Processes out-of-date sources again at a quiet hour (default 03:00)
Max parallel processing	2	How many sources are processed at once — lower it if your Mac gets hot
Local LLM timeout	10 min	Per-call ceiling for Ollama/MLX (big models on small Macs can be slow)
Delete wiki with source	off	When you delete a source, also delete the wiki page created from it

The Edit config.json button at the bottom opens the raw config for anything without a dedicated control (Notion row caps, extra auto-sync connectors, duplicate-merge thresholds, …).

Voice

Settings → Voice configures local transcription for audio dropped into raw/inbox/audio/ (see Voice memos):

Engine — faster-whisper (default), openai-whisper, or whisper.cpp.
Model — Whisper size (default base; bigger = slower + more accurate).

The Whisper CLI itself isn't bundled — the card shows which binary it expects. Transcription is entirely local; there is no cloud fallback.

MCP servers

Settings → MCP Servers connects Anthracite to external MCP servers as content sources: add a server (command + args), Test it, then Pull to ingest its documents into the vault. This is a power-user path — nothing here is required for normal use.

Developer settings — ingest token & webhook

Settings → Developer is where you get the credentials the capture APIs need. Two cards:

Local ingest API

Endpoint — the full http://127.0.0.1:<port>/v1/ingest URL (copy button).
Port — the bound port on its own, for clipper config.
Bearer token — show/hide + copy. Send it as Authorization: Bearer <token>.
Regenerate token — rotates the token (invalidates the old one; update any clipper/scripts that used it).

The token is stored in the vault's .anthracite/state.json and never leaves your machine. Each vault has its own token.

Webhook ingest

Webhook URL — the full http://127.0.0.1:<port>/v1/webhook/<secret> URL (show/hide + copy). The secret rides in the path, so senders that can't set headers (Apple Shortcuts, IFTTT) still work.
Regenerate webhook URL — rotates the secret.

Both are used in Capturing content, which has full curl examples.

Because the backend binds loopback, the webhook works for senders on this Mac out of the box; remote senders need a tunnel (a future Anthracite Cloud feature).

Connections

The Connections section holds the per-connector setup for reMarkable, Readwise, Notion, Apple Notes, Folder watch, and Obsidian (ingest-only). Step-by-step setup for each lives in Capturing content.

Connectors with accounts (reMarkable, Readwise, Notion) store their tokens in the macOS keychain — disconnecting removes the token. Each connector has its own Auto-sync toggle (5/10/15/30/60 minute interval); everything is manual-sync by default.