TeiNam1k downloadsAgentic LLM sidebar with four backends (AWS Bedrock, Google Gemini, OpenAI, Ollama). Includes chat, Graph RAG vault search, Second Brain layer, to-do management, web clipper, and MCP server integration.
This documentation describes 0.8.0. See the changelog for version history.
An AI assistant sidebar plugin for Obsidian with multi-provider backend support — AWS Bedrock, Google Gemini, OpenAI, and Ollama.
Note on command names: command palette entries, notices, the status bar, and tool results follow the UI language you pick in settings (English, 한국어, 日本語). Obsidian caches the palette at load time, so restart the app after switching languages to see the new names.
=, !=, >, >=, <, <=, and ~ (substring). Dot notation reaches nested keys, so project.status = active works. Invalid filters are reported rather than silently dropped, so a bad condition never masquerades as a whole-vault search.txt, .csv, .json, and .html files in your vault are indexed and searchable alongside notes, not just read once and discarded.base file with four views (decision ledger, open questions, outdated knowledge, review queue) from your vault data. Files you wrote yourself are never overwritten#heading anchors are checked against the index; unresolved citations are flagged under the answer.md and .txt work on all four backends; each sends up to its first 8,000 characters, and a cut is marked in the prompt and on the file chip. Images work on all four backends; PDFs on Bedrock and Gemini; Office documents on Bedrock. Unsupported combinations are refused with the list of backends that can handle the format, rather than dropped silentlyobsidian-markdown, obsidian-bases, json-canvas, korean-writing, business-english-writing, second-brainSupports desktop, phones, and tablets with Obsidian 1.13.2 or later.
Obsidian 1.13+ settings search can find the settings shown for the current backend. Earlier Obsidian versions retain the settings screen.
https://github.com/teinam/obsidian-agent-llmsmain.js, styles.css, manifest.json from the latest Release.obsidian/plugins/agent-llms/Version 0.4.0 changes the plugin ID from ai-assistant to agent-llms. The old ID was already claimed by an unrelated plugin in Obsidian's community plugin registry (qgrail/obsidian-ai-assistant), so Obsidian's updater treated this plugin's folder as that one and overwrote it. Nobody else has registered the new ID.
data.json), vault index, chat history, sessions, MCP config, and credentials are automatically copied on first launch. Your backend choice, models, region, Second Brain settings, and custom skills are all preserved. The old files remain in place, so rolling back to a previous version works seamlessly..obsidian/plugins/ai-assistant/ — if the overwriting update already ran, that folder holds an unrelated plugin's code rather than this one. The old data files (.ai-assistant-*.json, .bedrock-assistant-*.json) are no longer used either; you may delete them manually if vault size is a concern. The index file can be tens of MB due to embeddings.If you were using the kiro-edition (Assistant Kiro), the same migration applies. That edition was merged into main in 0.3.0, and .assistant-kiro-*.json data is also migrated automatically.
Settings → Agent LLMs → AI Backend:
The sidebar icon, model list, and branding update dynamically when you switch.
Backend support policy: This plugin relies on embedding APIs for Graph RAG vault search, so only providers with embedding endpoints are supported. The Anthropic direct API is excluded because it offers no embedding endpoint — use the Bedrock backend to access Claude models.
Bedrock: Set the AWS Region and configure authentication. On PC the plugin tries Roles Anywhere profile → SSO profile → Bedrock API key in that order (enter ~/.aws/config profile names); mobile uses only the Bedrock API key (AWS Console → Bedrock → API keys). If a profile works, the API key can stay unselected. The plugin does not refresh the API key automatically; replace the secret's value when it expires.
Key issuance, model access, and using multiple machines: Bedrock setup guide
Only the profiles and key entered in settings are used. The plugin never falls back to the AWS SDK default credential chain (
[default]profile, environment variables).
Required IAM permissions:
bedrock:InvokeModelWithResponseStreambedrock:InvokeModelbedrock:ListFoundationModelsbedrock:ListInferenceProfilesbedrock:CallWithBearerToken (API key only)Gemini: Select or create a secret that holds your API key from Google AI Studio.
OpenAI: Select or create a secret that holds your API key. To use an OpenAI-compatible endpoint, set the base URL including /v1; leave it empty for the official API.
Ollama: Enter the server base URL, or leave it empty to use http://localhost:11434. No API key is needed.
Key storage: API keys are kept in Obsidian's keychain (Settings → Keychain), on desktop and mobile alike. Each key field lets you pick or create a secret, and the vault's
data.jsonstores only the secret's name. Secrets are stored per vault and per device, so create a secret with the same name on each device and vault you use.
Keys saved by earlier versions are moved into the keychain the first time a vault loads: the desktop credential file (agent-llms-credentials.json under Electron's userData), the single mobile SecretStorage entry, and any key left in data.json. The desktop file is shared by every vault on that computer, so it is left in place; a vault that has moved its keys does not read it again. If the OS keychain is locked, the move is retried on the next start. Mobile preserves unreadable desktop ciphertext in data.json for desktop migration and never uses it as an API key.
Use the ribbon icon or run Open assistant from the command palette. Chat opens in the right sidebar on both desktop and mobile.
Click 🔍 in the chat header to index notes for semantic search, or run Index vault. Indexing is required for Graph RAG search and for the Second Brain tools that search the vault. emerge needs the index too, because it enumerates index entries. architect and update_index read the vault file list directly and work without an index.
On desktop, press Enter to send. On mobile, Enter inserts a newline; use the send button or Ctrl/Cmd+Enter to send. The AI responds in real-time streaming. Attach notes for context using the toolbar buttons:
.md, .txt, …), images, and PDFs from your computer via file picker, drag-and-drop onto the input area, or clipboard pasteThe web search toggle (globe icon) in the input toolbar only turns on if a search MCP (fetch, exa, or brave) is configured, or if you are on the Gemini backend, which has native Google Search grounding. Otherwise clicking it shows a notice and the toggle stays off.
Mobile does not run local MCP servers, so this toggle is available only for Gemini's native web search.
Settings → Agent LLMs → Generation Settings → Reasoning Effort sets how much reasoning the model does.
The current value is shown next to the model name at the bottom of the chat (for example, Opus 5.5 · high). Type /effort high in chat to change it, or /effort alone to see the current value and the levels the model accepts. The command is not sent to the model or saved in the conversation, and the change applies from the next request.
Allowed values depend on the selected provider and model (for example, Anthropic models on Bedrock accept xhigh and max; Gemini Pro models accept only low and high). The setting is only shown for models that support reasoning effort, and requests to models that do not support it fall back to the provider's default sampling behavior. If you switch to a model that does not allow your saved value, it is clamped to the nearest allowed level.
Notes are split into chunks and embedded, then a search walks outlinks and backlinks
from the best matches to pull in related neighbours. Start indexing from the search icon
in the sidebar header, or the Index vault command. Edited files are re-indexed automatically.
Details: Graph RAG & Second Brain
A layer that creates and maintains wiki notes grounded in your existing notes. It is off by default — enable it explicitly under Settings → Second Brain.
<!-- @generated:KEY --> markers, so regenerating keeps any notes you wrote yourself in the same file.Details: Graph RAG & Second Brain
Click the globe icon (🌐) in the action toolbar above the chat input → enter a URL. The page is fetched, translated (if needed), and summarized as a markdown note.
The generated frontmatter has four fields: source (the URL), created (the date), type: web-clip, and tags: [web-clip].
{{date}} / {{prevDate}} variables01. Projects, 02. Areas, 03. Resources, 04. ArchivesSettings → MCP Servers → Edit Config:
{
"mcpServers": {
"fetch": {
"command": "docker",
"args": ["run", "-i", "--rm", "mcp/fetch"]
}
}
}
Both uvx (Python) and docker are supported.
Declare API keys, proxies, DOCKER_HOST, and other server-specific values in that server's env; the full parent environment is no longer inherited. These values are stored in the vault's .obsidian/plugins/agent-llms/mcp.json, so manage that file's sync scope when it contains secrets. This is separate from local encrypted storage of AI backend keys.
Enable Confirm note changes and MCP tools to review the tool name and input before every MCP call (off by default). Stop all also cancels servers that are still initializing.
This plugin makes network requests to the following external services:
bedrock-runtime.us-east-1.amazonaws.com).generativelanguage.googleapis.com for chat, embedding, and model listing.https://api.openai.com/v1 for chat, embedding, and model listing, or to the OpenAI-compatible base URL you configure.http://localhost:11434), which is local unless you point it elsewhere.cdn.buymeacoffee.com.No data is sent to any third-party analytics or tracking services.
Process execution, filesystem access outside the vault, and MCP environment variables apply only on desktop. API keys are kept in Obsidian's keychain on every platform; vault and clipboard access are shared. MCP servers have their own access privileges.
child_process.spawn) — Saved MCP commands run on startup or reconnection with shell: false. The plugin does not insert a shell, but it does not sandbox the command or scripts supplied as arguments. No server configuration means no server process. A direct child that remains alive three seconds after stopping receives a force-kill signal.fs) — Earlier versions kept AI backend keys encrypted in agent-llms-credentials.json under Electron's userData directory. The plugin reads that file once per vault to move the keys into Obsidian's keychain, and does not modify or delete it. Credential files of legacy plugin IDs are copied within the same directory first. MCP executable lookup uses Node's spawn and PATH.PATH, HOME, USERPROFILE, APPDATA, LOCALAPPDATA, SYSTEMROOT, SYSTEMDRIVE, COMSPEC, PATHEXT, TMPDIR, TMP, TEMP, LANG, LC_ALL, and LC_CTYPE are inherited by default. Per-server env overrides are then applied. Other tokens and runtime options are not passed automatically. A server controls how it uses or transmits values explicitly provided to it.See the mobile support guide (Korean) for mobile behavior and validation, and the 0.7.14 review record (Korean) for earlier editing and attachment changes.
Release assets from 0.7.6 onward carry GitHub artifact attestations, so you can confirm they were built from this repository:
gh attestation verify main.js --repo TeiNam/obsidian-agent-llms
Chat, vault search, note editing, Graph RAG, Second Brain, and to-do management share the same code on desktop and mobile. Local MCP server execution and MCP settings are desktop-only.
For Ollama on mobile, configure a server address reachable from that device; localhost refers to the phone or tablet itself. Keep the app in the foreground for long indexing or agent tasks. See the device setup and testing guide (Korean).