MCP App Store
by KeenEthics
Guru icon

Guru MCP Server for Claude and ChatGPT

By Guru

Overview

Guru's MCP Server connects Anthropic's Claude to your company's trusted knowledge layer, so every AI answer is grounded in verified, permission-aware information—not outdated docs or hallucinations. It links Claude to the same governed context your teams use, ensuring consistent, cited, and explainable responses. With built-in verification and lineage, the Guru MCP Server makes it easy for Claude to deliver answers you can trust across every workflow.

How to connect

  1. Claude: Open the Guru listing in the Claude connectors directory and choose Connect. Claude listing Guru exposes 6 tools here.
  2. ChatGPT: Open the Guru listing in the ChatGPT app directory and choose Connect. ChatGPT listing Guru exposes 38 tools here.

Tools

fetch

ChatGPT
Fetch content by ID

guru_add_draft_collaborator

ChatGPT
Add a collaborator to a card draft in Guru. Provide either an email address for a user collaborator or a groupId for a group collaborator, but not both.

guru_answer_generation

ChatGPT
Generate direct answers to questions using the knowledge base. Use when users ask specific questions ('What is X?', 'How do I Y?'). Provides comprehensive answers, not just documents. If answer generation takes too long, this will execute asynchronously: the response includes a status field. If status is COMPLETE the answer is ready and included in the response. If status is PENDING the answer is still being generated — call the guru_get_pending_answer tool with the returned chatThreadId and turnId to retrieve it, repeating with a delay until status is COMPLETE or ERRORED.

guru_apply_draft_edits

ChatGPT
Edit a CARD draft by applying ops (this is for card drafts — the kind opened in the card editor — not page drafts). FIRST call guru_read_draft_for_edit, then call this with your ops — they anchor to that read automatically. Read again IMMEDIATELY before EVERY edit — never reuse an alias from an earlier edit or turn, as the document may have changed and the aliases would be stale. Ops merge into the draft instead of overwriting it. The response returns the merged document — check your change landed. If you get a stale-read error, read again and rebuild your ops. OPS FORMAT: A JSON array of edit ops. Each op targets a node by its alias — the alias field the read put on every node, at ANY depth (top-level nodes, and nested ones like a table cell or list item). Copy the alias exactly; never guess or compute it. To change content inside a table/list, target the specific cell/item by its alias — you do NOT rebuild the whole container. Shapes: {"op":"insertNode","before":"<alias>","node":<node>} (or "after":"<alias>" — inserts a sibling next to that node; to add at the END of the document, use "after":"<alias of the last TOP-LEVEL node in the read>" — not a nested child inside a table/list/callout, even if that nested node's alias appears last in the read; to add a content block at the end of a callout/collapsible, use "after" the last child inside it; to add a structural row/item INTO a table or list, e.g. a new row at the bottom of a table, use "after" the last row/item in that container — but if the table/list has zero rows/items (e.g. all were deleted in a prior edit), there is no row/item alias left to anchor after: instead use replaceNodeContent at the table's/list's OWN alias with "children" set to the new row/item node(s), to seed the first row/item back into the emptied container) | {"op":"replaceNodeContent","at":"<alias>","children":[<leaf/inline nodes>]} | {"op":"setNodeProps","at":"<alias>","props":{...}} | {"op":"deleteNode","at":"<alias>"} | {"op":"addEntity","key":"<fresh uuid>","entity":{"url":"...","name":"...","alt":"..."}}. When the document or a container exposes only a lone empty spacer (an empty draft body, or an empty callout/collapsible/table cell), author into it with replaceNodeContent on that spacer — or insertNode "before" it — NEVER insertNode "after" it, which strands the spacer as a leading blank line above your content. replaceNodeContent edits a node's interior (does not recreate it); use deleteNode only to remove a node. NODE SHAPES (build whole nodes for insertNode and for children): {"type":"title","children":[{"text":"..."}]} is the document title — always the first node, exactly one, plain text only; rename it with replaceNodeContent at its alias, and never insert or delete a title; paragraph {"type":"p","children":[<text>]} (use "p", not "paragraph"); {"type":"heading","level":1|2|3,"children":[<text>]}; {"type":"block-quote","children":[<text>]}; {"type":"list","listType":"ul"|"ol","children":[{"type":"list-item","children":[<text>]}]} (whenever content is a series of related items/points/steps — even if the user did not say "list", and especially items you wrote as bullets in your own reply — emit ONE list node: strip the bullet marker and put each item's text in a list-item. Never append "- "/"1." paragraphs. list-item holds text/inlines directly, do NOT wrap them in a p; a sub-list is a nested list placed as a SIBLING of the list-item, in the parent list's children (not inside a list-item — that placement isn't valid and normalization moves the sub-list out to a sibling); it renders indented under whichever list-item immediately precedes it, so place it right after that item, e.g. {"type":"list","listType":"ul","children":[{"type":"list-item","children":[{"text":"parent"}]},{"type":"list","listType":"ul","children":[{"type":"list-item","children":[{"text":"child"}]}]}]}); {"type":"code-block","language":"javascript","children":[{"type":"code-line","children":[{"text":"..."}]}]}; {"type":"markdown","children":[{"type":"markdown-line","children":[{"text":"<one source line>"}]}, ..…

guru_archive_card

ChatGPT
Archive a card. Archiving is a soft delete: the card disappears from search and folders but can be restored from the Guru web app's Archive view. Cards synced from external sources cannot be archived. In the response, 'archived' is false when the card was already archived.

guru_bulk_get_card_comments

ChatGPT
Get comment threads across all cards visible to the current user, most recently active first. Each result is a top-level comment plus all of its replies. activeAfter/activeBefore filter on the thread's last activity — the later of the comment's creation date and its newest reply's creation date — so a thread is returned when a reply falls in the range even if the top-level comment predates it. Results are ordered by that same last-activity date, descending. Status filtering (OPEN or RESOLVED) applies to the top-level comment only; replies have no status. Returns up to 25 threads per call; pass the returned nextPagingToken as pagingToken for the next page.

guru_bulk_get_cards

ChatGPT
Get full content and metadata for multiple cards in a single call (maximum of 50 card ids per call). Each returned card includes its id (always the canonical lowercase UUID), title, content (Markdown by default, or the stored HTML when contentFormat is HTML), url, verification state, and collection/folder info. Ids for cards that don't exist or that the user cannot access are skipped and returned (as canonical lowercase UUIDs) in 'notFoundOrInaccessibleCardIds' instead of failing the call. For a single card lookup (or to resolve a short id/slug from a card URL to a canonical id), use guru_get_card_by_id instead.

guru_create_attachment_upload_url

ChatGPT
Get a short-lived url for uploading ONE local file (e.g. an image the user attached) to Guru, to place it in a card or card draft. The upload responds with JSON whose link is the Guru-hosted url of the file: use it as the entity url in guru_apply_draft_edits, or as the src of an <img> in guru_update_card content. Only use it when the user asks to put a file or image into a card. Upload the file yourself: send method to uploadUrl as multipart/form-data with the file in the fileField part, e.g. curl -X POST -F "file=@<path-to-file>" "<uploadUrl>". Never base64 the file into a tool call. The url works for one upload and expires in 5 minutes, so request a new one per file. If the upload can't connect, your environment is probably blocking the url's host; tell the user an admin needs to allow it.

guru_create_draft

ChatGPT
Create a Draft of a Card in Guru. This tool only creates the draft shell — it does not accept content. Pass cardId to draft an edit of an existing published card; omit it to create a new standalone card. To add or change content afterward, call guru_read_draft_for_edit to read the draft's current content and get the ops format, then guru_apply_draft_edits to write content into it.

guru_create_folder

ChatGPT
Create a new folder in a collection. The folder is added at the top level of the collection unless a parentFolderId is provided. Folders can be nested at most 3 levels deep. Use the guru_list_folders tool to find the parentFolderId of an existing folder to nest under.

guru_delete_folder

ChatGPT
Delete a folder. The cards and subfolders inside the deleted folder are NOT deleted — they move up into the folder's parent (or the collection's top level). A collection's home (root) folder cannot be deleted. In the response, 'deleted' is false when no such folder exists (for example, it was already deleted).

guru_discard_draft

ChatGPT
Discard a draft, permanently deleting its unpublished changes. This cannot be undone — confirm the right draft with guru_get_draft before discarding. The published card the draft belongs to (if any) is unaffected; a standalone draft is deleted outright. Requires the draft's id (UUID) — find it with guru_list_drafts if you only have a card id or a content reference. Fails if the draft does not exist or the user cannot access it.

guru_get_card_by_id

ChatGPT
Get full card content, as Markdown by default. Pass contentFormat HTML for the card's stored HTML, which is what guru_update_card takes, so edit-then-save flows should read HTML. Accepts either the canonical card id (UUID) or the short id/slug from a card URL, and returns the card's canonical 'id' — use it to resolve a URL/short id before calling tools that need the canonical id (e.g. guru_list_drafts).

guru_get_dataset

ChatGPT
Get a dataset's fields, with the ID, name and type of each, its allowedValues if the values are fixed, its capabilities (SELECT, FILTER, GROUP_BY, ORDER_BY, COUNT, AGGREGATE) and its filter operators. ALL means every one. A content field also has a defaultLength, and a maxLength if its length has a limit. The dataset and a field can have notes on how queries on them behave. Call this before guru_query_dataset, which accepts only these field IDs, copied exactly. Some are upper snake case and some are UUIDs. Never re-case, reformat or invent a field ID, or send a display name in its place.

guru_get_document

ChatGPT
Get the content of a document by type and ID.

guru_get_draft

ChatGPT
Get a draft's title and current content in HTML format, including standalone drafts not linked to any published card. Requires the draft's id (UUID) — find it with guru_list_drafts if you only have a card id or a content reference. This is a plain read: to edit the draft use guru_read_draft_for_edit and guru_apply_draft_edits, and to throw its unpublished changes away use guru_discard_draft. In the response, cardId is null for a standalone draft.

guru_get_pending_answer

ChatGPT
Retrieve the result of a previously-initiated async answer. Call this with the chatThreadId and turnId returned by the answer generation tool when status is PENDING. Repeat with a delay between requests until status is COMPLETE or ERRORED.

guru_list_collections

ChatGPT
List every collection the user can access. Each collection includes its collectionId, name, description, type, and whether it is read-only. Use this to look up the collectionId for a collection by name before using other tools that operate on a collection.

guru_list_datasets

ChatGPT
List all available datasets, with the ID, name and description of each. Call guru_get_dataset for a dataset's fields before you query it with guru_query_dataset. Copy a dataset ID exactly as returned. Never build one from a prefix and another ID.

guru_list_drafts

ChatGPT
List the drafts the user can edit, most recently modified first, including standalone drafts not linked to any published card. Each draft's id can be passed to guru_read_draft_for_edit and guru_apply_draft_edits. Results are paginated: when the response includes a nextPagingToken, pass it back as 'pagingToken' to get the next page; only a null nextPagingToken means there are no more pages. Pass cardId to instead find the draft of an existing published card (a card has at most one); that returns an empty list if the card has no draft the user can access — note that guru_create_draft with a cardId fails if a draft already exists.

guru_list_folder_items

ChatGPT
List the items that live directly inside a folder, in folder order: cards (type 'card') and subfolders (type 'folder'). Find a folderId via guru_list_folders (or guru_list_collections then guru_list_folders). To descend into a subfolder entry, call this tool again with the subfolder's id; to read a card entry's content, use guru_get_card_by_id. Results are paginated: when the response includes a nextPagingToken, pass it back as 'pagingToken' to get the next page. A page may contain fewer items than the page size (or none); only a null nextPagingToken means there are no more pages — keep paging until nextPagingToken is null.

guru_list_folders

ChatGPT
List the folders the user can access, optionally scoped to a single collection. Each folder includes its folderId, title, collectionId, the number of cards directly inside it, and whether it is a collection's home (root) folder. Results are paginated: when the response includes a nextPagingToken, pass it back as 'pagingToken' to get the next page. A page may contain fewer items than the page size (or none); only a null nextPagingToken means there are no more pages — keep paging until nextPagingToken is null. Typical chaining: guru_list_collections to pick a collectionId, this tool to find a folderId, then guru_list_folder_items to see the cards and subfolders inside that folder (pass a home folder's folderId to list a collection's top-level content).

guru_list_groups

ChatGPT
List user groups on the team, optionally filtered by a search term. Each group includes its groupId and name. Use this to look up the groupId for a group by name before adding or removing it as a draft collaborator.

guru_list_knowledge_agents

ChatGPT
Get available knowledge agents and their IDs. Always call this first before using guru_search_documents or guru_answer_generation tools, as both require an agentId parameter. Each agent has different expertise areas and data sources configured. If the user doesn't specify which agent to use then find the appropriate agent for a specific domain or use case based on name. For search and answer generation, do not default to 'Guru' agent unless there are no high confidence matches.

guru_move_card

ChatGPT
Move a card from one folder to another within the same collection. Cards can live in more than one folder; when this card does, pass sourceFolderId to say which occurrence to move. To move a card to a DIFFERENT collection, use the guru_move_card_to_collection tool instead.

guru_move_card_to_collection

ChatGPT
Move a card to a different collection, optionally into a specific folder there. Requires permission to move the card out of its collection and to publish in the destination collection. A private card can only be moved by its author, and cards synced from external sources cannot be moved. Moving a card to the collection it is already in is a no-op.

guru_move_folder

ChatGPT
Move a folder under a different parent folder within the same collection, or to the collection's top level when parentFolderId is omitted. The folder keeps all of its contents. Folders can be nested at most 3 levels deep, a folder cannot be moved into itself or its own subfolders, and a collection's home (root) folder cannot be moved. To move a folder to a DIFFERENT collection, use the guru_move_folder_to_collection tool instead.

guru_move_folder_to_collection

ChatGPT
Move a folder to a different collection, bringing all of its cards and subfolders along. Requires write access to the folder and permission to create folders in the destination collection. Folders can be nested at most 3 levels deep, and a collection's home (root) folder cannot be moved. Moving a folder to the collection it is already in is a no-op.

guru_query_dataset

ChatGPT
Query a dataset: list the records that match a filter, or aggregate them, with grouping, ordering and paging. Use a search tool for text search. Call guru_list_datasets, then guru_get_dataset for the dataset's fields, then query. Copy dataset and field IDs exactly as returned. A dataset has only the fields it lists. Read the notes of the dataset and of each field you use, which say how queries on them behave. Easy to get wrong: - With only field select items, the query lists records, one row each. With an aggregation select item, it returns one row per group of its field select items. groupByFieldIds and having need an aggregation, and groupByFieldIds must equal the field select items or be omitted. - Every select item, filter and having condition needs its 'type', also when nested in a combo condition. - Count rows with COUNT on fieldId ''. '' is not a field and works only with COUNT, with distinct false. - Use a field only as its capabilities and operators allow, where ALL means every one. COUNT covers COUNT and COUNT DISTINCT. AGGREGATE covers MIN, MAX, SUM and AVG, and SUM and AVG need a numeric field. A distinct query can select only fields with GROUP_BY. - There is no IN, CONTAINS, LIKE or NULL operator. Match several values with a combo OR of EQUAL conditions, and test for a missing value with EQUAL null. - A filter on an ID field needs a well-formed ID, or the query fails. An ID which does not exist, or which you cannot access, matches nothing, with no error. - A content field, such as card or source record content, lists a defaultLength. A field select item's maxLength sets how many characters to return, and a row's truncatedColumns lists the 0-based indexes of its values that were cut. - A response holds at most 100 rows, or 10 when a select item asks a content field for more than its defaultLength and the field can be longer. Send its pagingToken to get the next page, until no token comes back. If truncated is true, the result reached the most rows the dataset can return and more can match: narrow the filter to get the rest.

guru_read_draft_for_edit

ChatGPT
Read a CARD draft's live content so you can edit it with guru_apply_draft_edits (for card drafts — the kind opened in the card editor — not page drafts). Returns the draft's nodes, each carrying an alias. Target ops at those aliases — copy them exactly; apply_draft_edits anchors to this read automatically. Edits merge into the draft instead of overwriting it. Always read immediately before editing. Also returns the draft's entities map (image/file metadata keyed by entityKey); new image/file nodes may reference an existing key without a fresh addEntity.

guru_remove_draft_collaborator

ChatGPT
Remove a collaborator from a card draft in Guru. Provide either an email address for a user collaborator or a groupId for a group collaborator, but not both.

guru_search_documents

ChatGPT
Find and retrieve specific documents, sources or cards. Use when users want to discover existing content ('find me docs on X', 'do we have cards about Y'). Returns documents for browsing, not answers.

guru_update_card

ChatGPT
Update the content of an existing card. Use when users want to modify existing card content.

guru_update_collection

ChatGPT
Update the name, description, color, or emoji of an existing collection. Only the provided fields are changed; omitted fields keep their current values. Use the guru_list_collections tool to find the collectionId of the collection to update.

guru_update_folder

ChatGPT
Update the title or description of an existing folder. Only the provided fields are changed; omitted fields keep their current values. A collection's home (root) folder cannot be updated. Use the guru_list_folders tool to find the folderId of the folder to update.

guru_verify_claim

ChatGPT
Check factual claims against the knowledge base. Use when a user wants to know whether statements, a draft, or an answer are consistent with Guru knowledge ('is it true that X?', 'fact-check this'). Returns one entry per claim with a status: SUPPORTED (a document states it), NOT_SUPPORTED (a document states something that conflicts with it), or INSUFFICIENT_INFO (the knowledge base does not settle it; this does not mean the claim is false). A NOT_SUPPORTED claim comes with a short summary of what Guru says ("Guru says up to $3,000"), a one-line suggestion that does not restate the claim ("To match Guru, use $3,000."), a longer explanation, and the refuting sources. When possible it also has conflictingText, the exact part of the claim that is wrong, and correctedText, the text that replaces it. Each source has a quote, a url, an iconUrl to show beside it, and its verification: verified, verificationState, lastVerifiedBy and lastVerifiedAt. An unverified source is still evidence; show the reader that it is not verified. SUPPORTED and INSUFFICIENT_INFO claims have no explanation or sources unless verboseMode is true. Each claim should be one fact; split compound statements before calling. At most 10 claims per call.

guru_verify_text

ChatGPT
Find the factual claims in a block of text and check them against the knowledge base. Use when a user wants a draft, a message or an answer fact-checked against Guru knowledge and has not split it into claims. Pass the whole text unchanged. Returns findings, sorted by position. Each finding has text, the part of the input it is about, and start and end, 0-based character offsets into the input (end exclusive), so that text is input.substring(start, end). These three are null, and the finding comes after the placed ones, when the claim could not be found in the input because it contains masked data. Each finding also has claim, the self-contained claim that was checked, and a status: SUPPORTED, NOT_SUPPORTED (a document states something that conflicts with it) or INSUFFICIENT_INFO (the knowledge base does not settle it; this does not mean the claim is false). A NOT_SUPPORTED finding has a short summary of what Guru says, a one-line suggestion, an explanation and the refuting sources. When text is exactly the wrong part, it also has correctedText, the text that replaces it. Each source has a quote, a url, an iconUrl and its verification: verified, verificationState, lastVerifiedBy and lastVerifiedAt. An unverified source is still evidence; show the reader that it is not verified. Only NOT_SUPPORTED findings are returned unless verboseMode is true. At most 10 claims are checked per call. Use verify_claim instead when the claims are already separated.

answer_generation

Claude

create_draft

Claude

get_card_by_id

Claude

list_knowledge_agents

Claude

search_documents

Claude

update_card

Claude

Similar Apps

Listed on MCP App Store

Listed on MCP App Store

Add this badge to your site or README. It links to this page.

HTML

<a href="https://mcp-app-store.keenethics-labs.com/app/guru"><img src="https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg" alt="Listed on MCP App Store" width="220" height="44"></a>

Markdown

[![Listed on MCP App Store](https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg)](https://mcp-app-store.keenethics-labs.com/app/guru)