Every AI Entry Point

DocGraph’s AI shows up in ten different places. They share a model and a privacy posture but differ in what they read, whether they touch your files, and which formats they work on. This page is the map; each row links to where it’s explained properly.

Everything here except MCP is off until you enable AI and requires a chat model (set one up); the MCP server has its own toggle and runs against the connected agent’s model, not one of yours. Nothing on this page contacts a network service unless you have deliberately set one up — a cloud provider, a remote Ollama, or web search.

Where AI appears

Entry point	Where	Reads	Writes	Formats
Chat panel	Right sidebar	Vault or open file	Only via agent actions, confirmed	All
Summarize	Toolbar ✦ AI menu	Open file	No	All readable
Proofread	Toolbar ✦ AI menu	Open file	On Fix	HTML, Markdown
Beautify	Toolbar ✦ AI menu (not in <>Source)	Open file	Immediately	HTML
/ AI commands	Slash menu	Open file (Ask AI: your vault)	Yes — Ask AI and Write with AI as a draft you accept	HTML only
Selection AI	Selection toolbar ✨	The selection	Yes	HTML, Markdown
Typing suggestions	Inline, as you type	Current paragraph	On Tab	HTML only
File right-click	File tree	That file	Tags only	Text-extractable
Folder right-click	File tree	Up to 12 notes in the folder	Only when you click Create note from this or Create table	Text-extractable
MCP	External agent	Whatever it asks for	Confirmed, or auto-approved	All

The two summarize buttons

There are two, and they are not the same thing:

• ✦ AI menu → Summarize this file summarizes the file you have open, into a popover. Nothing is written.
• / → Summarize inserts that summary into the page as a blockquote.

Both reuse the same cached result, so the second is instant after the first. A cached summary is also what powers the TL;DR bar that can appear above a page in reading view — it never runs the model just because you opened a file, so no TL;DR means no summary has been made yet.

File right-click actions

Right-clicking a file in the tree offers three AI items, on any file whose text can be extracted — including PDFs, spreadsheets and Word documents:

• Summarize — a modal with the summary. No write.
• Suggest tags — proposes tags as clickable chips; clicking one applies it. This writes to DocGraph’s tag index, not into the document.
• Ask AI about this — opens the chat panel with that file as context.

Which model runs what

Two models are configured in Settings → AI → Chat & index:

• The chat model answers in the chat panel.
• The utility model runs the background jobs: summaries, proofreading, tag suggestions, commit messages, and the quick rewrites chat makes of your question before searching — and the typing suggestions in HTML notes (grey text after the cursor, Tab to accept). With no utility model set, these use the chat model.

Splitting them lets you pair a large chat model with a fast utility one. The trade-off is real and worth knowing: a small utility model makes Proofread nearly useless — see the warning on the editor AI page.

None of these features need a model with tool-calling or thinking support. Those capabilities only matter for the chat panel’s agent actions.

What ⌘Z can undo

Action	Undoable
Proofread Fix	Yes
Accepting a typing suggestion	Yes
Rewrite / Shorten / Fix grammar	Yes
/summarize, /continue	Yes
/write Accept or Delete	No
Beautify	No — Revert bar only, and only until you navigate away
A chat agent change	Not with ⌘Z — use ↩ Undo last AI change under the answer (most changes)
Anything an MCP agent writes	Not with ⌘Z — it goes on the same undo stack as the chat agent’s changes

For what neither can reach, version history is the real safety net.

Markdown gaps

The .md surface has less AI than HTML does. The full comparison is on the editor AI page; the short version is that the toolbar buttons and selection actions work, and the slash-menu AI group and typing suggestions do not exist there at all.

Privacy in one place

• AI is off by default and nothing is downloaded or contacted until you turn it on.
• With a model on this computer — the built-in runtime, Apple’s on-device model or a local Ollama — no vault content leaves your machine for the AI. The status pill in the top-right reads 100% local when that’s true.
• Web search is off by default; when on, each query is shown to you before it is sent, and a SearXNG server on your own computer keeps it local.
• Pointing DocGraph at a remote Ollama requires a separate, explicit confirmation, and secret-detection then inspects what would be sent.
• A cloud provider (OpenAI, Anthropic, OpenRouter, a custom endpoint) is off by default: you save a key, allow it in a system dialog, and pick one of its models. The badge then reads Cloud AI · {provider}, and secrets are blocked from the prompt by default.
• Every AI run is recorded in Settings → Privacy → Activity & audit — model, task, token counts, and which notes were placed in context. Not a transcript.
• Two paths are outside that boundary by design and say so: external CLI agents run under their own credentials, and an MCP client relays whatever it reads to its own provider.

Every AI tool

In Agent Mode the assistant works through tools, and an MCP agent gets the same set. There are 87: 43 read-only tools that run without asking, and 44 write tools that stop and ask you first (Asks in the tables) — unless you choose Approve all this session in chat, or an MCP client’s auto-approve covers them. With Agent Mode off, only read-only tools are offered — the write tools are not merely blocked, they are absent — and a model usually starts with a short menu of the common ones, asking for the rest when it needs them.

Two more tools exist only in the in-app assistant: web_search, when web search is on, and propose_web_search, which offers to turn it on when it is off. They are never exposed over MCP, so an MCP agent gets the same 87.

Find and read notes

Tool	Kind	What it does
search_vault	Reads	Search for relevant passages by keywords, meaning and link structure, with optional filters such as tag, person and date. Returns passages with their file paths and the links between them.
find_notes	Reads	List the notes a request means (“find the note about X”), with the reason each was picked; the best one opens when it is clearly ahead.
read_file	Reads	Read a file’s text, a page at a time for long files.
read_spreadsheet	Reads	Read exact cells and rows from a spreadsheet (.xlsx, .xls, .ods, .csv), by sheet, with paging.
list_files	Reads	List files and folders by exact path, folder prefix, filename fragment, or files with no readable text only.
get_outline	Reads	A file’s heading tree, to find the right section before reading a long document.
get_note_context	Reads	One note’s properties, tags, outline, the notes linking to it (each with the sentence around the link), the notes it links to and the links among those notes, in one call — what an agent reads before editing a note.
open_file	Reads	Show a file to you at the end of a turn. Navigation only; nothing is changed.
list_locked	Reads	Names the encrypted files that exist. Their contents are never readable until you unlock them.
docgraph_help	Reads	Answer a question about DocGraph itself — features, settings, shortcuts — from the copy of this user guide bundled with the app, citing the section it used.

Create and edit notes

Tool	Kind	What it does
create_file	Asks	Create a new HTML note. Never overwrites an existing file.
edit_file	Asks	Find-and-replace edit, on HTML and Markdown notes. The card shows what is removed and what is put in.
append_to_file	Asks	Append an HTML fragment to the end of a page.
copy_file	Asks	Duplicate a file byte-for-byte, including images and PDFs.
rename_file	Asks	Rename or move a file and rewrite the links that pointed at it.
delete_file	Asks	Move one or more files to the OS trash, with an undo entry each.
list_templates	Reads	List the notes in the Templates folder.
create_from_template	Asks	Create a note from a template, filling {{title}} and the date tokens.
daily_note	Asks	Return the path of a day’s daily note, creating it from the template if it doesn’t exist. Never makes a second note for the same day.
capture	Asks	Add a note or a todo to the capture inbox note.
add_task	Asks	Add a todo in one step — its text, the people it is for and its due or start date — to the capture inbox, today’s daily note, the open page or a named note. Dates in your own words are read by DocGraph on your clock and shown on the card.
get_authoring_guide	Reads	The page-conventions reference the AI follows when it builds pages.
validate_page	Reads	Lint an HTML page it wrote: page skeleton, table, query and chart marker ids, in-vault links, scripts.

Links and the graph

Tool	Kind	What it does
list_backlinks	Reads	Notes that link to a note, with depth and PageRank.
get_outbound_links	Reads	Notes a note links to, with depth and PageRank.
graph_neighbors	Reads	Notes linked to or from a note, in both directions.
run_graph_query	Reads	Walk the link graph from a note by direction, depth and link type (rel).
shortest_path	Reads	The shortest link chain between two notes (up to 6 hops), with each hop’s link type.
find_orphans	Reads	Notes no other note links to.
find_broken_links	Reads	Links that point at files that don’t exist.
set_link_rel	Asks	Set or clear the typed link (rel) on the links from one HTML page to a target. HTML pages only.

Tags, properties and people

Tool	Kind	What it does
get_tags	Reads	The tags on one file, both yours and auto-generated.
list_tags	Reads	Every tag in the vault with the number of files carrying it.
list_pages_by_tag	Reads	Every file carrying a tag.
apply_tags	Asks	Apply up to 12 tags to a file. Written to DocGraph’s tag index, not into the document.
query_properties	Reads	Find pages by their page properties (key/value metadata).
set_property	Asks	Set one page property.
delete_property	Asks	Remove a page property.
get_ontology	Reads	The vault’s declared note types, entity kinds, link types and typed property keys.
list_people	Reads	Every @name the vault mentions, most-mentioned first.
mentions_of	Reads	Notes that mention a person with @name.

Tasks and block properties

Tool	Kind	What it does
list_tasks	Reads	The vault’s todos, filterable by status, assignee and dates. Each is addressed as <path>#<ordinal>.
set_task	Asks	Tick a task off, or un-tick it.
set_task_props	Asks	Set a task’s assignees, assigner and start and due dates. Omitted properties are cleared. Dates in your own words are read on the card, as for add_task.
set_block_props	Asks	Set or remove properties on HTML lines or todos. One call can change many files, so it always asks, even with Approve all this session on.

Folders and the file tree

Tool	Kind	What it does
create_folder	Asks	Create a folder.
move_folder	Asks	Move a folder and rewrite the links into it.
delete_folder	Asks	Delete a folder and everything in it, to the OS trash. Undoable in-app only up to 100 files / 2 MB.
set_folder_icon	Asks	Set a folder’s or a page’s icon and colour.
clear_folder_icon	Asks	Remove a folder’s or a page’s icon.
get_tree_order	Reads	The tree’s manual sibling order.
set_tree_order	Asks	Set the manual order of a folder’s children.
set_pinned_files	Asks	Pin or unpin files at the top of the tree.

Tables and queries

Tool	Kind	What it does
list_tables	Reads	List the vault’s data tables.
get_table_schema	Reads	A table’s columns and types.
query_table	Reads	Filter, sort and page a table’s rows.
search_tables	Reads	Keyword search inside table cells.
get_related	Reads	Follow a relation column to the row it points at in another table.
list_saved_queries	Reads	List the saved queries.
run_saved_query	Reads	Run a saved query.
query_source	Reads	Query the vault’s own data — files, tasks, tags, people, dates, links, properties — without a saved query.
create_table	Asks	Create a data table.
rename_table	Asks	Rename a table.
delete_table	Asks	Delete a table with all its rows and columns. Undoable with Undo last AI change.
add_column	Asks	Add a column.
update_column	Asks	Rename a column or change its type; cell values are kept.
insert_row	Asks	Add a row.
update_row	Asks	Change a row.
delete_row	Asks	Delete a row.
set_table_sort	Asks	Set a table’s display sort, like clicking its column headers.
save_query	Asks	Save a named query over a table.
delete_query	Asks	Delete a saved query.
import_spreadsheet_as_table	Asks	Import a .xlsx or .csv as one editable table per sheet, plus a page that embeds them. The source file is copied, not changed.

Charts and whiteboards

Tool	Kind	What it does
list_charts	Reads	List the saved charts.
read_chart	Reads	Read the categories and values a chart plots right now, to check it before citing it.
create_chart	Asks	Create a bar, line, area, pie, scatter or Gantt chart from a table or query, or update an existing one by id.
delete_chart	Asks	Delete a chart.
read_board	Reads	A whiteboard’s nodes, edges and frames (no positions). Called before edit_board.
create_board	Asks	Create a whiteboard or diagram from a description of nodes and edges.
edit_board	Asks	Add, change or connect nodes on an existing board.

Version history

Tool	Kind	What it does
git_status	Reads	Uncommitted changes in the vault.
git_history	Reads	Past snapshots of the vault, or of one file.
git_snapshot	Asks	Take a local version-history snapshot. Never pushes.

Scanned documents

Tool	Kind	What it does
make_searchable	Asks	Make scanned PDFs or images searchable. Their text is read on this computer, in the background.
remove_from_search	Asks	Take PDFs or images out of search and delete the text read from them.

Memory

Tool	Kind	What it does
read_memory	Reads	Read Memory/agent.md, the lines remembered across chats.
propose_memory	Asks	Propose one line (up to 300 characters) to remember. You approve it first; it is never auto-approved.