Writing Assistant Chat

AI writing assistant for Obsidian with a unified chat and editing surface. Connects to local or cloud LLM providers. Features vault-wide RAG retrieval, knowledge graph, agentic tool use, note context, and reusable prompt commands.

Desktop only.

Features

One conversation, two ways to work

• Ask before edits, The default. Write operations follow the per-operation policy in settings, which normally asks you to review changes before they are applied.
• Edit automatically, Applies proposed changes without per-operation review. Use it for trusted sessions where hands-off editing is intentional.

Multi-provider support

Connect to one or more LLM providers:

• LM Studio, Local inference via OpenAI-compatible API. No cloud, no API keys, no data leaving your machine.
• Anthropic, Claude models with native API support and prompt caching.
• OpenAI, GPT models via the OpenAI API (or any OpenAI-compatible endpoint).
• Claude Code, Anthropic's agent harness via the local claude CLI. Runs its own tool loop over the vault toolstack and uses your existing Claude Code login, so no API key is needed.

Switch between providers and model profiles from the chat panel.

Agentic tool use

When enabled, the model can use tools across multiple reasoning rounds.

Every operation that changes your vault is routed through an approval gateway. By default each kind (create, overwrite, move, trash, create folder, in-document edit) is set to Ask, so nothing is applied without your review. In settings you can set any kind to Auto-apply or Deny (the tool is removed entirely so the model is never offered it).

Vault-wide retrieval (RAG)

Semantic search over your entire vault using local or cloud embeddings. Configurable chunk size, overlap, similarity threshold, and metadata enrichment (tags, folder paths, wikilinks).

Knowledge graph

LLM-powered entity and relationship extraction from your vault. Enables entity-based retrieval and graph-aware ranking of search results.

Note context

The active note is automatically available to conversations, so the model writes with awareness of your current document. Configurable context size.

Prompt commands

Reusable prompt templates (e.g. "Tighten dialogue", "Expand this scene") that appear as buttons in the chat panel and in the editor right-click context menu. Supports {{selection}} and {{noteText}} placeholders.

Streaming and message management

• Real-time streaming responses.
• Message version history with regeneration.
• Token usage tracking and cost estimation.
• Inline message editing.
• Chat history with conversation switching.
• Draft auto-save.

Model profiles

Save multiple configurations per provider, each with its own system prompt, temperature, max tokens, top-p, top-k, and reasoning level. Switch profiles from the chat panel.

Requirements

• Obsidian v1.0.0 or later (desktop only; tested on recent 1.12.x releases)
• At least one LLM provider:
  • LM Studio, lmstudio.ai, running locally with at least one model loaded
  • Anthropic, An API key from console.anthropic.com
  • OpenAI, An API key from platform.openai.com
  • Claude Code, The Claude Code CLI installed and signed in (no API key needed, will use your existing Subscription)

Installation

From community plugins

1. Open Settings > Community plugins > Browse
2. Search for Writing Assistant Chat
3. Click Install, then Enable

Beta via BRAT

1. Install the BRAT plugin from community plugins
2. Open Settings > BRAT > Add Beta plugin
3. Enter Resolve-public/writing-assistant-chat and click Add Plugin
4. Enable Writing Assistant Chat in Settings > Community plugins

Getting started

1. Open Settings > Writing Assistant Chat
2. Choose a provider and configure it:
   • LM Studio, Start the local server (default http://localhost:1234) and the plugin will discover loaded models
   • Anthropic / OpenAI, Enter your API key
   • Claude Code, Make sure the claude CLI is installed and signed in (set its path only if it isn't on your PATH)
3. Add a model profile with your preferred system prompt, temperature, and token limit
4. Click the chat icon in the ribbon, or run the command Open writing assistant chat
5. Start writing

Network and privacy

Remote services

When using cloud providers, the plugin sends your messages (and any note context you include) to the provider's API:

When using Local providers exclusively, no data leaves your machine.

Data handling

• API keys are stored locally in Obsidian's plugin data file and are only sent to their respective provider.
• Conversations, RAG embeddings, and knowledge graph data are stored locally on your device.
• No telemetry, analytics, or tracking. The plugin makes no network requests beyond what is required to communicate with your chosen provider.
• No account required. LM Studio needs no account; Anthropic and OpenAI require their own API accounts.
• Cloud usage is billed to you. The plugin is free. When you use a cloud provider, Anthropic and OpenAI charge metered usage against your own API key (this includes OpenAI embeddings if you select an OpenAI embedding model for RAG), and Claude Code draws on your existing Claude subscription. Local providers such as LM Studio are free to run.

External tools and files outside your vault

The Claude Code provider is the only part of the plugin that reaches outside your vault:

• It launches an external program. When you use this provider, the plugin runs the claude command-line tool as a subprocess, with its working directory set to your vault. To find the executable it searches the directories on your system PATH, and on Windows it reads the npm shim files (claude.cmd / claude.bat) to resolve the real program. It only ever runs a claude binary that you have already installed; it never downloads, installs, or updates one.
• The CLI keeps its own files outside the vault. The claude tool stores its login, configuration, and per-conversation session files in its own home-directory folder (~/.claude), outside your Obsidian vault. The plugin relies on these to resume a conversation after a restart. These files are created and managed by the Claude Code CLI itself.

Every other tool the plugin offers the model (read, edit, create, move, trash) is path-restricted to your vault.

Support

This plugin and all of its features are, and will always be, free. If it helped you get closer to achieving your creative goals, you can support this project in the following ways:

• Buy Me a Coffee

You can also find these links in Settings > Writing Assistant Chat > General.

Contributing

See CONTRIBUTING.md for setup instructions, project structure, coding standards, and the development workflow.

License

MIT, see LICENSE for details.

The distributed build bundles third-party components under their own licenses, including one proprietary Anthropic component that is not covered by this plugin's MIT license. See THIRD-PARTY-NOTICES.md for the full list and terms.