com.local-mcp/local-mcp

Deletes a file or empty folder from OneDrive.

Returns metadata for a file or folder: size, modification date, type, and extension. Faster than listing the parent directory when you only need info about one item.

Lists files and folders in a OneDrive path. Use onedrive_root to find valid paths. Returns up to `limit` entries (default 1000, max 5000); large folders are truncated with a note — narrow the path for more specific results.

Moves or renames a file/folder within OneDrive.

Reads a text file from your OneDrive synced folder. Supports .txt, .md, .csv, .json, .xml, .log and several code file types. Auto-detects UTF-8, falls back to Latin-1/Windows-1252 for legacy files (common in Latin American banking .TXT padrones). For files elsewhere on this Mac, use file_read.

Lists all mounted OneDrive directories on this Mac.

Searches for files by name in a OneDrive directory (recursive). Returns up to max_results matches (default 50); raise max_results or narrow the root for more.

Restricts LMCP's OneDrive access to a specific folder. Once set, all OneDrive tools (read, write, list, search, delete, move) only work inside the allowed folder. Pass an empty folder to remove the restriction. Changes take effect immediately.

Writes text content to a file in OneDrive.

Checks which email accounts are configured in Microsoft Outlook and compares them with Mail.app. If Outlook has accounts not in Mail.app, guides the user to add them so all email tools work seamlessly.

Reads and extracts text from a PDF document (.pdf file).

Reports the TCC permission state (screen recording, accessibility, microphone) this app needs to capture the screen and drive other apps' UI. Call it before a capture/automation run and surface the grant hints instead of failing mid-sequence. Screen Recording / Accessibility are granted in System Settings (not a JIT dialog); the URLs open the exact pane.

Creates a PowerPoint presentation (.pptx) at `path` from an array of slides, each {title, bullets:[…]}. Requires confirm=true — called without it, returns a preview of the deck instead of writing the file. The path must be somewhere Local MCP can write; Desktop/Documents/Downloads may need a one-time Files-and-Folders grant (System Settings → Privacy & Security → Files and Folders). Returns {created, path, slides}.

Reads slide text content from a PowerPoint presentation (.pptx file).

Use this when the user wants the full content of an email that lives in the Mac's Apple Mail (message ID from list_emails/search_emails). For a Microsoft 365 message ID from m365_list_emails, use m365_read_email. Pass account= (and mailbox= if known, both from list_emails/search_emails) so the lookup targets one account instead of scanning all of them. Call sequentially, not in parallel — concurrent calls serialize behind Mail.app's JXA lock and later calls will time out. Performance: body fetch is the primary latency source (avg 20s on slow IMAP). Pass include_body=false to skip it and get metadata-only (fast). Pass max_body_chars=N to cap the body at N chars after HTML stripping (default 30000; 0=unlimited). Response includes body_fetch_ms when fetch took >2s, body_omitted=true when skipped, body_truncated_at=N when cut. When a body isn't cached on this Mac, read_email returns metadata with body_omitted=true and body_omit_reason="not_downloaded" (iCloud/IMAP optimized storage) rather than making Mail fetch it (that can be slow and tie Mail up). If the user wants it anyway, retry with force_download=true to have Mail pull the body over IMAP now and return it (waits up to ~60s). Off by default; ignored while Mail is in a cooldown.

Reads messages from an iMessage conversation by chat ID or contact name.

Reads the full content of a note by name or ID.

Use this when the user wants to remove one of THEIR saved recipes/skills (the manifests under ~/.local/share/local-mcp/recipes). Destructive with a preview gate: the first call (without confirm) shows what would be deleted; call again with confirm=true to actually delete. Bundled starter recipes can't be deleted. To modify a recipe instead, recipe_save with the same name overwrites it (upsert).

Browses installable community SKILLS — ready-made LMCP workflows other people published (a morning brief, inbox triage, a weekly report). A user would browse them to find a ready-made workflow for a repeatable multi-app task instead of building it from scratch. Returns a list of {id, title, category, description, steps, votes}; install one with recipe_install(id).

Exports a saved SKILL (recipe) as a single portable token the user can send to someone else — paste it in a message, email, or doc. The recipient installs it with recipe_import and runs it with recipe_run. A user would export a skill to share it with a teammate (a handy brief, a report, a workflow). Returns {name, skill_token} plus the readable manifest.

Returns the full manifest of a recipe by name. recipe_not_found if unknown.

Installs a SKILL someone shared with you — pass the skill_token from their recipe_export (or a raw recipe manifest JSON). Saves it to this Mac so recipe_run can use it. Safe: importing only stores the skill; when it's later run, any state-changing step (send/write/delete) previews first and needs confirmation. If a skill with the same name already exists, the import is saved under a non-colliding name. Returns {name, imported}.

Installs a community SKILL by id (from recipe_discover) onto this Mac so recipe_run can use it. Safe: installing only stores the skill; when it's later run, any state-changing step (send/write/delete) previews first and needs confirmation. If a skill with the same name already exists, it's saved under a non-colliding name. Returns {name, installed}.

Lists the user's reusable SKILLS — saved recipes (an ordered sequence of LMCP tool calls with parameters), plus bundled ones — each runnable with recipe_run. Skills turn a repeated LMCP workflow into one reusable command. A user would list them to find an existing skill for a task rather than rebuilding it from scratch. Returns each skill's name, description, and steps.

Executes a recipe end to end: binds params, runs each step's tool in order via the registry, persists the run (see recipe_runs), and returns each step's result plus any markers_path. Recipes with state-changing steps (write/send/delete) PREVIEW first — call again with confirm:true to execute; read-only recipes run immediately. A step that errors stops the run and is reported.

Shows the history of past recipe runs and their results (recorded by recipe_run), so you can reuse, compare, or debug an automation. Pass `name` for one recipe's runs, or omit for a compact history across all recipes. Pass `run_id` (with `name`) to get that run in full detail. Newest first.

Saves a reusable SKILL — a named recipe (an ordered sequence of LMCP tool calls with parameters) — to this Mac so the user can re-run it anytime with recipe_run. A user would save one to turn a multi-step LMCP workflow they repeat (a morning brief, inbox triage, a weekly report, a data pull) into a single reusable command. Saved skills can be shared with other people via recipe_export. The manifest must have a name and a non-empty steps array. Returns {name}.

Drops a named marker into the active recording's timeline. t_ms is elapsed ms since recording start. Provide bounds (global points, top-left) to zoom toward an element, or omit for full-frame. note becomes a caption source. Returns no_active_session if nothing is recording.

Renames an existing Apple Reminders list. Pass the current list name (or list_id from get_reminder_folders) and new_name. Requires confirm=true.

Use this when the user wants to reply to an email that lives in the Mac's Apple Mail (message ID from list_emails/search_emails). Supports plain text or HTML body. For a Microsoft 365 message ID from m365_list_emails, use m365_reply_email. Pass `account` (from list_emails/search_emails results) to skip scanning other accounts and avoid timeouts on multi-account Macs.

Send an ANONYMOUS, content-free signal when an LMCP tool fails, returns nothing useful, the user seems frustrated, or you could not accomplish what they asked. Helps the LMCP team find and fix the roughest spots. Send ONLY the category + the tool name — NEVER the user's request, message/email content, account names, or any personal data. No confirmation needed: this is anonymous (categories only) and respects the user's opt-out.

Sends a problem report, feature request, or integration request to the LMCP team — for when a user wants to report a bug, ask for a new capability, or request support for an app LMCP doesn't cover yet. Without confirm=true it returns a preview of the anonymous payload that would be sent (version, OS, and permission status only — no personal data); with confirm=true it submits and returns a case_id. type='problem' (default) reports a bug, type='feature' requests a new capability, type='integration' requests an unsupported app.

Submits a feature request to the LMCP team — a new capability, a tool that doesn't exist yet, or an app/integration the user wishes LMCP supported. A user would invoke this when they want to ask the team for something LMCP can't do yet. Without confirm=true it returns a preview; with confirm=true it submits. The request includes the machine ID and (if set) the account email so the team can follow up — it is not anonymous.

Runs a fast health check of all LMCP integrations on this machine. Shows what works, what doesn't, and how to fix it. Optionally submits a report to the LMCP team.

Clicks the first element matching a CSS selector in the current Safari tab. Returns the tag name and visible text of the clicked element so you can confirm the right thing was hit. Pass `wait_for_navigation: true` to wait up to 3 seconds for the page to load after the click (useful when clicking links or buttons that trigger navigation).

Fills multiple form fields in one shot. Pass fields as a JSON object mapping CSS selector to value.

Navigates the current Safari tab back to the previous page.

Lists or searches the user's Safari browsing history (local SQLite, read-only — no page is opened). Optional `query` matches the URL or page title (case-insensitive substring). Returns url, title, visit_count and last_visit (ISO), newest first. Requires Full Disk Access. For Chrome use chrome_history.

Lists every open tab across all Safari windows with title, URL, and whether it is active.

Navigates Safari to a URL. Pass new_tab=true to open in a new tab.

Runs document.querySelectorAll in the current Safari tab and returns a compact summary of each match.

Reads the rendered text content of a Safari tab. Identify the tab either by `url_match` (substring match against URL; first hit wins) or by `window_index` + `tab_index` (from safari_list_tabs). Text is capped at `max_bytes` (default 100 KB). Pass `include_html: true` to also get the raw HTML source. Pass `include_links: true` to extract all links with their href and text (useful for following navigation in SPAs like dashboards).

Searches the rendered text of every open Safari tab for a substring. Returns each matching tab with the surrounding snippet. Useful for 'do I have a tab open with X?' questions across dozens of tabs.

Reports whether Safari is ready for interactive tools (safari_click, safari_type, safari_evaluate_js). Returns setup instructions if JavaScript from Apple Events is not enabled.

Sets the value of an input/textarea matching a CSS selector and fires input/change events.

Polls the current Safari tab until a CSS `selector` appears, or until `text_match` appears (anywhere on the page, or inside `selector` if you also pass one). Pass at least one of `selector`/`text_match`. Useful after safari_click to wait for the next page or a modal to render.

Saves an attachment from an email to disk. Pass account= (and mailbox= if known, both from list_emails/search_emails) so the lookup targets one account instead of scanning all of them.

Begins a screen recording (ScreenCaptureKit) of a display, window, or region. Single active session in v1 — a second start returns already_recording. Returns a session_id used by record_marker and screen_record_stop. Requires Screen Recording permission; without it returns an explicit permission_required error, never a silent no-op.

Reports whether a recording is active, with the session_id, elapsed_ms, output path, and marker_count.

Stops the active recording, finalizes the .mov, and writes the marker timeline JSON (§6) next to it. Returns the video path, duration, resolution, marker_count and markers_path. Returns no_active_session if nothing is recording.