Eraser MCP

PREFERRED for creating a diagram from a natural-language prompt — Eraser's AI picks the diagram type, generates the DSL, and renders it. DO NOT pre-classify the user's request into a diagram type yourself. Pass `text` only (plus destination params) and LEAVE `diagramType` UNSET unless the user EXPLICITLY named a type (e.g. 'make this a sequence diagram', 'use a cloud architecture diagram'). Eraser's render service has type-specific heuristics and will pick a better type from the prompt than your guess will, especially on ambiguous asks like 'show how X works' or 'diagram the architecture'. BEFORE calling this tool, if the user has not already specified the destination, ASK them these questions (skip any the user has already answered): 1. 'Should I create a new file for this diagram, or add it to an existing file?' — if existing, ask which fileId (or use one they just referenced). 2. If NEW file: 'Should the file be private (just for you) or shared with your team in a folder?' 3. If shared in a folder: 'Which folder?' (or offer to place it at the team root). Do NOT call list_folders or get_folder to invent a destination yourself — ask the user. Pick destination params based on the answers: - Existing file → pass targetFileId - New private file → omit targetFileId, omit folderId, do NOT set isTeamFile - New shared file at team root → omit targetFileId, omit folderId, set isTeamFile: true - New shared file in a folder → omit targetFileId, pass folderId (this implies shared) Returns the new diagram's id in the `diagrams[].id` field so a follow-up update_diagram/manually_update_diagram call does NOT need a list_diagrams round-trip. Prefer over manually_create_diagram unless you already have hand-written DSL/JSON.

PREFERRED for populating a document from a natural-language prompt — Eraser's AI generates the markdown. Only callable on a file with an empty document body; if the file already has content, this returns an error and you should call update_document (for natural-language edits) or manually_update_document (to replace verbatim) instead. Prefer over manually_create_document unless you have pre-written markdown to insert verbatim.

Create an empty file, or duplicate an existing file when sourceFileId is provided. Never populates content from a prompt — use create_document/create_diagram for AI generation. BEFORE calling this tool, if the user has not specified destination, ASK: 1. 'Should the file be private (just for you) or shared with your team in a folder?' 2. If shared in a folder: 'Which folder?' (or offer to place it at the team root). Do NOT call list_folders to invent a destination — ask the user. Pick params: - Private file → omit folderId, do NOT set isTeamFile - Shared at team root → omit folderId, set isTeamFile: true - Shared in a folder → pass folderId (this implies shared)

Create a new folder.

Create a new preset (the team-level container for AI styling: templates, references, and rules). After creating, the typical next steps to make the preset useful are: 1. Add example files as templates/references via `add_or_remove_template_or_reference` (or `create_template_or_reference` to spin up a fresh template file from a prompt). Templates anchor the visual style; references anchor terminology and concepts. 2. Set the preset's rules via `update_rules` (e.g. 'always use dark backgrounds', 'prefer sequence diagrams for auth flows'). 3. Optionally mark the preset as the team default by calling `update_preset` with `isDefault: true`. This tool only creates the empty preset shell; an agent that stops here will produce a preset with no styling power. Chain the steps above unless the user explicitly only wanted the shell.

Create a new template (style anchor) or reference (terminology/concept anchor) file and, when `presetId` is provided, attach it to that preset in a single call. Auto-publishes the file's first version. Templates and references are ALWAYS team-scoped resources under AI Presets — there is no concept of a private/personal template or reference. DO NOT ask the user 'team template or private template?' — that is not a real choice. The only sharing question is which preset to attach it to. Typical use: - User says 'make this a template under ' → pass `presetId` so the file is attached immediately. - User says 'make a template from this file' without naming a preset → ASK them which preset before calling (or list_presets to show options). DO NOT silently omit `presetId`; an unattached template is invisible to AI generation and almost never what the user wanted. - User says 'duplicate this file as a template' → pass `sourceFileId` (and `presetId`). When `presetId` is omitted the response includes a warning that the file is detached; treat that as a signal to follow up with `add_or_remove_template_or_reference` rather than leaving it dangling.

Delete a diagram from a file.

Clear a file's document body to an empty markdown document.

Archive a file.

Delete a folder. Rejects when the folder is not empty.

Delete a preset. Rejects when the preset has any rules, templates, or references.

Render a canvas diagram to PNG or JPEG and return a temporary image URL. Tell the user to download it from the returned imageUrl.

Export a file's markdown document body as a downloadable artifact.

Returns the Eraser file URL for PDF export. NOTE: programmatic PDF export is not yet available via MCP — this returns a link for the user to export from the Eraser app's export menu, not a downloadable PDF. For a diagram image, use export_diagram; for the document body as markdown, use export_document.

Fetch a diagram's metadata and DSL/JSON code. For freeform diagrams, set includeFreeformDefinition: true to get the full scene structure (elements, connections, titles). Need a PNG image of the diagram? Use export_diagram instead.

Fetch the full markdown body of a file's document.

Fetch a file's metadata, document outline (headers), and the list of diagrams in it.

Fetch a folder by id.

Fetch the current user, active team, and team memberships.

Fetch a preset including its rules, templates, and references.

Fetch a template/reference's metadata, document outline, and diagram list.

List the diagrams contained in a file.

List files in the team workspace, optionally scoped to a folder.

List folders. Pass `parentFolderId` to scope to direct children of a folder (or `null` for top-level only). Pass `nameContains` to resolve a folder the user names by string (e.g. 'the Engineering folder') without paging through the entire tree. Combining the two narrows further (e.g. top-level folders containing 'eng').

List the team's presets. Pass `nameContains` to resolve a preset the user names by string (e.g. 'the Marketing preset') without scanning the full list — much cheaper in context tokens.

List the teams the current user belongs to (OAuth only).

ADVANCED — most callers should use create_diagram instead. This tool writes a caller-supplied diagram definition VERBATIM into a new diagram; no AI runs. For DSL diagrams (flowchart-dsl, sequence-dsl, etc.) pass the DSL source as `code`. For freeform diagrams pass a freeform definition as `code` — either JSON ({"title"?: string, "elements": [ ... ]}) or the equivalent JSX; it is parsed and laid out through the freeform render pass (text measurement, connection routing, etc.). Use ONLY WHEN: the user gave you a literal diagram definition to paste, you are converting from a non-Eraser source (other tool's export, programmatic generation), or you need byte-exact output for a reason the AI path cannot satisfy. For any natural-language description of what the diagram should contain (e.g. 'a sequence diagram showing checkout'), use create_diagram — it picks the diagram type, generates the definition, and renders it for a fraction of the output tokens of composing it yourself.

ADVANCED — most callers should use create_document instead. This tool populates an empty file's document body with caller-supplied markdown VERBATIM; no AI runs and the bytes you pass are exactly what gets stored. Only callable on a file with an empty document body; if the file already has content, call manually_update_document (verbatim overwrite) or update_document (targeted AI edit) instead. USE ONLY WHEN: the user gave you literal markdown to paste, you are converting from another format (HTML, Confluence export, programmatic generation), or you need byte-exact output. For any natural-language description of what the document should say, use create_document.

ADVANCED — most callers should use update_diagram instead. This tool writes a diagram's complete DSL/JSON (or freeform edits) VERBATIM; no AI runs and the bytes you pass are exactly what gets stored. USE ONLY WHEN: - the user gave you literal DSL/JSON they want pasted as-is, - you are converting from a non-Eraser source (other tool's export, programmatic generation), - or you need byte-exact output for a reason an AI refinement cannot satisfy. DO NOT use this tool just because the current diagram code happens to be in your context (e.g. you just called create_diagram and want to apply a small refinement). For ANY user request phrased as a natural-language edit ('remove the X group', 'add a step', 'change the color', 'rename Y to Z') — even when you already have the full DSL in front of you — call update_diagram instead. Re-emitting the entire DSL here to apply a conceptual edit costs the user large amounts of output tokens, risks accidental edits to unrelated parts, and gives a worse UX than letting Eraser apply the targeted change.