Connect local MCP clients and Skills
Your goal
Allow other AI clients to read or edit notes with explicit permissions.
Before you start
Connect a local folder first. Practise in the demo vault if you prefer.
Steps
The existing workspace read/write switches include attachments. After restarting an updated backend, refresh both local clients and the LeafLoom ChatGPT plugin tools. vault_files provides search, metadata, original-byte reads, ZIP downloads and task status; vault_files_update provides chat/local uploads, move/rename, trash, organization, index maintenance and task control. Inspect a plan before confirming it with the current task revision. Uploads preserve duplicate names and return links without editing notes automatically. See the file tool contract.
Open Settings → AI connections → MCP access permissions. Enable local AI reading and copy the generated connection configuration. Its dedicated credential is shown once. Use http_headers.Authorization in Codex, headers.Authorization in Cursor, and type: http plus headers.Authorization in Claude Code. Regenerating the credential or disabling all its resource scopes revokes it; keep configurations private.
Enable Allow local AI to modify notes separately for direct saves. For ChatGPT, use the existing Tunnel connection and its separate read all notes and modify notes switches. Permission changes automatically reconnect the Tunnel; refresh the tool list in your client. Writing is off by default for both connections.
Full access covers all connected Markdown, TXT and HTML/HTM notes, including archives and notes outside RAG. It supports finding, reading, creating, precise edits, properties, moves, recoverable deletion, history, links, templates and daily notes. RAG still uses only selected eligible Markdown/TXT. HTML is read and edited as source. Existing unauthenticated loopback and legacy bearer connections retain the original five read-only capabilities; the old auth-token does not unlock full access.
Every mutation checks the workspace revision; existing notes also require their read-time version. Moves and deletes first return an impact preview; history and Trash can be previewed and restored. MCP reads saved files, preserves conflicting editor drafts and reports pending indexing separately from successful saves. The browser assistant keeps its proposal workflow for notes. Managed designs use design_document, design_operations and the design history, asset, library and export tools with the connection's read/write permissions.
Install LeafLoom skill from Settings → Skill, then use $leafloom in a local Codex chat. It connects directly to the running app for knowledge search and task operations. For MCP clients, see the tool contract.
On Windows, update the installed skill to use scripts/leafloom.ps1: it selects an available Python 3.11+ interpreter, skips Store aliases and sends PowerShell hashtables as UTF-8 JSON. Open a new chat after updating. See the platform-specific examples.
AI asset search and delivery
Before searching, enable Use for AI asset search in each desired attachment group. All groups start off; only the selected groups’ union enters keyword/semantic retrieval, and no selection returns scope_not_configured without a full-library fallback. This choice syncs to NAS with group metadata. Conflicted groups are excluded until explicitly resolved. Normal attachment browsing and note references remain available.
In Downloads → Download settings, set the Asset delivery directory and optionally enable semantic asset indexing with your prepared local model. AI can search existing local/NAS assets through vault_files actions search_assets, asset_details and asset_preview, then use vault_files_update.deliver_asset and poll delivery_status. Only completed deliveries return an absolute path on the identified service computer. Delivered files remain independently of cache cleanup; the calling AI handles project integration. Keyword search works without a model; package contents and visual similarity are not indexed.
What to expect
Full workspace access is independent of RAG scope; installing a Skill does not configure MCP.
If it does not work
Reload clients after updates or credential rotation and inspect the actual capabilities returned by status.