Notalith

Notalith is a platform-neutral Obsidian assistant powered by your own model-provider account. It provides a chat interface with tools to read Vault content and create or edit Markdown notes.

The plugin uses public Obsidian APIs and does not require Node.js, a local CLI, ACP, Electron APIs, or a locally spawned agent process at runtime.

Notalith is free and open source under the Apache 2.0 license. It has no paid features or Notalith account requirement.

Features

• Azure Foundry Responses, Chat Completions and Claude Messages, native Claude Messages and Gemini generateContent, and Chat Completions for DeepSeek, OpenAI, Grok, and OpenRouter.
• Multiple model profiles across providers, with a separate endpoint and API key for each provider.
• Streaming assistant responses and cancellable requests.
• Vault read tools and basic Markdown write tools with visible execution status.
• Local backlinks, outgoing and unresolved links, combined inline/frontmatter tag counts, and note outlines.
• Full Markdown/path search with tag, frontmatter, filesystem-time and folder filters, plus safe regular expressions and pagination.
• Import images and files from the device into the Vault using Obsidian's attachment location or a configured folder, without overwriting existing files.
• Current-note, editor-selection, note, image, and Office attachments.
• Vision input from Vault images and images embedded in Markdown notes when the selected model supports it.
• Local DOCX, PPTX, and XLSX text/table extraction.
• Obsidian SecretStorage for provider API keys.
• Optional image-generation tool shared by all chat providers, with independent Azure/OpenAI image configuration, Vault saving and previews.
• Desktop and mobile-compatible architecture based on public Obsidian APIs.

Model, message, and attachment IDs use native crypto.randomUUID() where available, with a crypto.getRandomValues() UUID fallback for mobile WebViews. Failures to add or save a model are reported in a notice rather than silently leaving the button unresponsive.

Supported content

Office files are parsed locally. Notalith sends extracted text and table data to the active provider rather than uploading the original Office file. Chat Completions and native Gemini do not accept the plugin's structured image tool result, so read_image is not exposed to those adapters. Manual and embedded-note images can be sent to vision-capable Azure deployments, deepseek-flash, and models with Image input enabled under Claude, OpenAI, Grok, Gemini, or OpenRouter.

Read-only tools

Directory tools return ordered items, total, offset, and nextOffset; follow nextOffset until it is null. list_directory lists immediate children, while get_directory_tree recursively returns a flat tree representation with Vault-relative paths and entry kinds. Existing list_files and list_notes retain their original non-paginated behavior.

Range tools use one-based inclusive startLine/endLine. Their offset and nextOffset are character offsets within the same selected range; repeat with unchanged line bounds to read all content, including unusually long lines. Null bounds select the whole file. Editor coordinates are instead zero-based line/ch. Missing notes, unavailable editors, invalid ranges, and unresolved anchors return explicit errors.

Knowledge relationships and enhanced search

Relationship tools use Obsidian's local metadata/link cache, so results reflect its latest indexing state. get_outgoing_links returns resolved targets; get_unresolved_links returns missing targets. Reference counts include repeated links and embeds. Tags combine frontmatter and inline occurrences, deduplicated within each note. An unindexed outline returns an explicit error; tag/search results include uncachedNoteCount to identify potentially incomplete metadata.

search_notes scans full Markdown content and paths, without an embedding API or a separate search index. Its filters are AND-combined: all requested tags (optional #, parent tags include subtags), all frontmatter conditions, a folder including descendants, and inclusive file creation/modification time bounds. An empty/null query returns filter matches. Frontmatter keys support dotted nesting; scalar equality is type-sensitive, contains checks an array member or substring, and gt/gte/lt/lte compare numbers. exists accepts a null value. Results are path-sorted, paginated, and include an excerpt and the one-based first content-match line (null for filter-only/path-only matches).

Dates are YYYY-MM-DD (UTC midnight) or ISO timestamps with a timezone. For an entire day, specify the end timestamp rather than that day's midnight. With regex=true, queries use RE2 multiline syntax; matching defaults to case-insensitive. Patterns are limited to 1,000 characters. Backreferences and lookarounds are unsupported and reported as errors. The browser-compatible RE2JS engine avoids catastrophic backtracking without Node.js or WebAssembly. The original search_vault remains available with its original keyword behavior.

All tool paths are validated as Vault-relative paths. Absolute paths, path traversal, and access to the Vault's Obsidian configuration directory are rejected.

Markdown write tools

The model can invoke these tools without a separate approval prompt. Replacement checks the latest file content using Obsidian's atomic Vault.process(); missing or repeated old text fails without changing the note. There is no whole-note overwrite, deletion, move, or binary-write tool. A stopped request prevents further tools but does not undo writes already completed. Review important notes or keep Vault backups before asking the model to edit them.

Requirements

• Obsidian 1.11.4 or newer.
• An API account for Azure Foundry, Anthropic (Claude), OpenAI, xAI (Grok), Google (Gemini), DeepSeek, or OpenRouter.
• At least one configured model with function calling for the plugin's Vault tools. Image requests additionally require a model with image input.
• The appropriate endpoint, model or deployment name, and provider API key.

You provide your own account and API key. Notalith does not charge you, but your model provider may; check its pricing and usage limits before sending requests. The plugin connects directly to your configured endpoint, without a Notalith-operated intermediary.

For Azure, the deployment name, not the underlying model family name, is sent as the selected API's model value. For every other provider, enter an exact model ID available to your account. OpenRouter IDs generally include a provider prefix such as openai/gpt-4.1-mini. The example IDs shown in settings are placeholders, not a model catalog.

Azure Foundry protocols

Choose API protocol for each Azure deployment: OpenAI Responses (the unchanged default for existing profiles), OpenAI Chat Completions, or Claude Messages. Deployment aliases never determine the protocol, and errors do not trigger automatic protocol switching. Changing protocol or endpoint starts a new conversation.

For Claude, a standard Azure resource endpoint is mapped to https://<resource-name>.services.ai.azure.com/anthropic/v1; requests use x-api-key and anthropic-version: 2023-06-01. Custom gateways require an explicit Endpoint override. Overrides are API base URLs without the final /messages, /responses or /chat/completions route, and use the Azure connection's saved key. OpenAI protocols use /openai/v1 base URLs. Enable Image input only for vision-capable Chat Completions or Claude deployments. Chat Completions does not support image tool results.

Claude thinking and redacted-thinking blocks are retained with their signatures for subsequent turns and tool results, but are not shown as assistant text. The implementation follows the official Claude on Foundry API examples without a runtime SDK dependency.

Model router

Add your router's exact deployment name (for example model-router), select OpenAI Chat Completions, and set Endpoint override to https://<resource-name>.services.ai.azure.com/openai/v1. This uses the v1 Chat Completions API without an api-version parameter and reuses the Azure connection's saved key. The router's selected underlying model can change between requests; the plugin continues to send the router deployment name and the full conversation/tool history, not the selected model's name.

For the tested resource, the openai.azure.com/openai/v1 endpoint returned deployment-not-found for the router, while the services.ai.azure.com v1 Chat Completions endpoint succeeded. The resource-level v1 Responses endpoint reported that the operation was unsupported. Support through Foundry project Responses endpoints is a separate API surface. The plugin does not silently switch endpoints or protocols. Routing modes and candidate model pools remain configured in Foundry, not in the plugin.

Manual installation

Notalith is not yet distributed through the Obsidian community plugin catalog.

1. Install dependencies and build:

   npm install
   npm run build
   

2. Create the plugin directory in your Vault:

   <vault>\.obsidian\plugins\notalith
   

3. Copy these files into that directory:

   manifest.json
   main.js
   styles.css
   

4. Open Settings → Community plugins, reload installed plugins, and enable Notalith.

For local development on Windows, the project directory can be linked directly:

New-Item -ItemType Junction `
  -Path "<vault>\.obsidian\plugins\notalith" `
  -Target "<project>\notalith"

Configuration

Open Settings → Notalith.

1. Choose a provider from the Provider dropdown. This only changes which settings are visible; it does not switch the active chat provider.
2. Enter your Azure OpenAI v1 endpoint, or keep the selected provider's default endpoint. Save its API key with Save key.
3. Select Add model and enter a display name and exact Azure deployment name or provider model ID. Several models can share a provider's endpoint and key.
4. For Claude, OpenAI, Grok, Gemini, or OpenRouter, turn on Image input only if that particular model supports images. Azure can send images to compatible deployments, and DeepSeek image input is limited to deepseek-flash.
5. Optionally select Test on the model. This sends a small live request to your provider, which may incur usage charges; it does not change the active model.
6. In the chat view, click Provider · model beside the send button. The menu groups models by provider; selecting a model switches to its provider and starts a new conversation. The model's Use model button in settings does the same.

Only profiles with a non-empty model ID appear in the chat menu; a saved key is still required to send a message. The settings dropdown's (active) marker shows which provider the current chat model uses. Azure response IDs and other providers' in-memory conversation histories are never reused across model switches. Existing Azure endpoints, deployments, active model selections, and saved keys migrate automatically.

API keys are stored in Obsidian SecretStorage; they are not written to plugin data.json, notes, or chat content. Provider reasoning metadata and Gemini thought signatures needed to continue tool calls are held only in memory, never shown in chat, and cleared with the conversation.

Usage

Automatic context compression

The plugin automatically checks estimated input tokens before every model request, including requests that follow a tool-result batch. Under Settings → Notalith → Context, the Context input token budget defaults to 32,000. Set it below your deployment's context window, leaving room for output/reasoning tokens. The plugin does not infer a model's window from its deployment name.

When the budget is exceeded, the selected chat model produces one bounded rolling text summary in an isolated request with no tools. Each update uses the previous summary plus only newly archived, completed interactions, not the entire original transcript. Recent interactions stay verbatim where they fit the retention budget; unfinished tool-call/result sequences are never split. The summary preserves goals, constraints, decisions, unfinished work and relevant paths. Image pixels and private reasoning are not sent to the summarizer, so old visual details are not guaranteed to survive compression.

The replacement targets roughly 60% of the input budget, with up to 8,000 estimated tokens of recent interactions and a summary capped at 2,000 tokens. System instructions, tools, new input and a safety allowance count toward the budget. Estimates account separately for non-ASCII text and image allowances and are calibrated conservatively with available provider input usage; they are not exact tokenizer counts or a guarantee against provider limits.

Compression updates model context without clearing the visible chat. Azure Responses starts a fresh response chain with the summary and retained items; this does not use a server-side compaction API. Other providers replace their local request history. The chat reports estimated before/after input size and summary-call usage when returned by the provider.

Summarization incurs an additional model call and can lose detail. It reduces repeated historical input in subsequent turns, not necessarily the total cost of a short conversation. Empty/oversized summaries, insufficient reduction, cancellation and concurrent resets never install a partial summary. An error stops the pending request instead of silently exceeding the budget or discarding history. If a single current input or unfinished interaction is too large, reduce its size or raise the budget; compression does not truncate it. Completed vault actions are not undone on failure.

Conversation context and summaries remain in memory only and reset with a new chat, model change or plugin reload. No manual compression action or session persistence is required.

Image generation tool

Under Settings → Notalith → Image generation, select an Azure Foundry or OpenAI connection and enter the image deployment/model ID (for example, gpt-image-2). Image configuration is separate from the chat model menu and reuses the selected connection's saved key. Enable Image generation to expose generate_image to the current chat provider. Disabled or incomplete configuration does not advertise the tool. An enabled/configured status is not a connectivity guarantee.

Azure uses the deployment-specific Images API with the configurable 2025-04-01-preview version by default. A standard /openai/v1 connection base URL is accepted and converted to the deployment route; a custom image endpoint override should be the Azure resource base URL. OpenAI uses its API base URL and /images/generations. The implementation follows the official Azure image-generation examples.

The first version supports text-to-image, one PNG per request, three preset sizes and low/medium/high quality. Choose parameters supported by your deployment; editing, masks and batch generation are not implemented. The Test button generates and discards one low-quality test image and can incur charges. No image request is made just by configuring or enabling the tool.

Ask any function-calling chat model to generate an image. The plugin saves it using the existing attachment folder/location rules and non-overwriting names, then returns only the saved path and embed link to the chat model, not Base64 image data. The chat displays an image card independently of the model's text. Use Copy embed link or Insert into note; insertion explicitly selects a Markdown note and appends an embed. Generation itself never changes a note. If explicitly requested, the chat agent can insert the returned embed using the existing Markdown write tools.

Generation requests are not automatically retried. After an image failure, further generation attempts in that chat turn are blocked to avoid accidental duplicate billing. If generation succeeds but saving fails, the image is kept in memory with Retry save and Discard actions in chat and settings. Retrying save does not call the image API again. Save or discard the pending image before another generation; reloading or unloading the plugin loses an unsaved image. Cancellation stops subsequent saves/edits where possible, but Obsidian's non-streaming request cannot cancel a server-side billed generation, and already saved images or completed note edits are not undone.

Open Notalith from the ribbon or run Notalith: Open chat from the command palette.

Examples:

Summarize the current note.

Search the vault for notes about Azure Foundry and cite their paths.

Describe test/diagram.png.

Read reports/quarterly.xlsx and summarize each worksheet.

Create notes/meeting.md with today's meeting outline.

Append a follow-up section to notes/meeting.md, then replace the exact text "Action: TBD" with "Action: review the draft".

Use the toolbar to attach:

• The current Markdown note.
• The current editor selection.
• A Vault note, image, DOCX, PPTX, or XLSX file.
• A supported plain-text file, or a reference to another Vault file.

Import files from your device

Click Import files into vault (upload icon) to open the device's file picker; select one or more images or files. Files are saved locally before being attached to the composer. The size limit is 25 MB per imported file. Images sent to the model still have a 10 MB limit. Unsupported formats, including PDFs and legacy Office, can be stored and linked but their contents are not extracted or sent to the model.

Imported attachment folder in settings overrides the destination with a Vault-relative folder. Leave it empty to follow Obsidian's attachment setting relative to the current Markdown note. Missing folders are created and duplicate filenames receive an available name; existing files are never overwritten. Invalid/protected paths and import failures are reported explicitly.

Use the link icon on an attachment chip to copy a link for the current note. Models can also call get_attachment_link. Links use Obsidian's configured Markdown/wikilink format. Importing or generating a link does not insert it into a note, and removing a chip does not delete the imported file. Import alone does not send file contents to a provider; sending the chat does. Downloading attachments from URLs is not implemented.

Privacy and security

• Vault content is read only when explicitly attached or requested through a model tool call.
• Tool activity, including write successes and failures, is shown in the chat UI; writes do not require approval.
• Notes and extracted Office content used in a request are sent to the selected provider endpoint.
• Images used in a request are Base64-encoded and sent only to a supported model on the selected provider.
• Office documents are limited to 25 MB.
• Images are limited to 10 MB.
• Attached notes and extracted documents are truncated according to the configured maximum character count.
• Imported plain-text attachments share that character limit; models can continue reading with read_text_file.
• Knowledge/search processing is local, but tool results and attached content used in a chat are sent to the chosen provider.
• No embeddings are generated and there is no automatic Vault upload.
• Vault content is treated as untrusted data in the system prompt.

Development

npm install
npm run dev

Production validation:

npm test
npm run build
npm run lint
npm run format:check

When Obsidian CLI is enabled:

obsidian plugin:reload id=notalith
obsidian dev:errors
obsidian dev:console level=error

Community release

Before submitting to the Obsidian Community directory:

1. Verify the plugin with configured models in Obsidian on desktop and mobile, and run the production validation commands above.
2. Keep package.json, manifest.json, and versions.json aligned on the release version. Publish a GitHub release tagged with the exact manifest.json version (without a v prefix) and upload main.js, manifest.json, and styles.css as release assets. main.js is generated by npm run build and is not committed to Git.
3. Submit the public GitHub repository through the Obsidian Community developer dashboard, select Optional payments because external model providers may charge users, and address any automated review feedback.

Current limitations

• Markdown writes are limited to creating, appending, and unique exact-text replacement; no whole-note replacement, delete, move, built-in backup, or undo.
• Chat Completions and native Gemini models cannot use read_image through tools. Azure and image-enabled Claude models can use it. Only deepseek-flash supports DeepSeek image attachments; Claude, OpenAI, Grok, Gemini, and OpenRouter require per-model image opt-in.
• No conversation persistence or session history.
• No PDF or legacy Office (.doc, .ppt, .xls) extraction.
• No OCR for scanned documents.
• No semantic/embedding search or URL-based attachment downloads.
• Office extraction focuses on readable text and table values; advanced formatting, charts, macros, comments, and embedded Office media are not preserved.
• Mobile compatibility is a design target and still requires device-level validation.

See DESIGN.md for the full product and architecture design, and ROADMAP.md for the Markdown editing status and future work.