VaultCortex
vault-cortex is a standalone MCP server that gives an AI agent direct read and write access to an Obsidian vault's markdown files, without Obsidian itself needing to be running. Search combines three techniques: FTS5 keyword matching for exact terms, vector embeddings for semantic similarity, and a cross-encoder reranking step for intent-heavy queries, with all models running locally rather than calling an external API.
The server adds a lightweight memory layer on top of the vault: an append-only folder (an "About Me" folder by default) where dated entries accumulate under topic headings, so corrections are added as new entries rather than overwriting history, and an agent can retrieve just the relevant entries instead of loading an entire file into context. It also understands two common Obsidian task-tracking conventions (the Tasks plugin's emoji syntax and Dataview's inline fields), letting an agent filter, complete, reprioritize, or move tasks in a single call. Non-text files are also made accessible: images pass through as images, PDFs as extracted or rendered text, and canvases as readable outlines.
Local setup takes about two minutes via a single init command with Docker and a vault folder. Remote deployment is available as one-click options on Render or Railway, or self-hosted, with OAuth 2.1 and PKCE securing MCP client access (6-hour access tokens, 60-day refresh tokens) or static bearer tokens for direct access. The ready-made remote image also bundles Obsidian's proprietary Sync service, which requires an active subscription to use.
vault-cortex is placed in Assess. It extends the same underlying capability as ClaudeObsidian (agent access to an Obsidian vault) with a broader feature set -- hybrid search, structured memory, and task management -- packaged as a self-hostable MCP server. It has not yet been trialed first-person.
