Connecting AI clients (MCP)
DataCNVS runs a Model Context Protocol (MCP) server, so you can work on your canvases from the AI tools you already use. Ask Claude or Claude Code to find an observation, build a multi-panel figure, convert an axis to different units or restyle a canvas, and it makes the edits in DataCNVS as you, where you and your collaborators see them live.
On this page
Connected AI clients build and edit the plot: panels, axes, series, labels and styling. They never receive the numbers in your observations. DataCNVS draws the figure from your data inside the app. How this works.
Before you start #
- A paid DataCNVS plan. MCP is included on every paid DataCNVS plan. On a workspace whose plan doesn't include it, every tool explains that and does nothing.
- MCP turned on for your workspace. Workspace admins control this in Settings → Workspace → General → AI → MCP / external AI clients. Turning it off also keeps AI clients away from that workspace's canvases and observations wherever they are shared.
- Your own permissions. The AI client acts as you. It can open exactly the canvases and observations you can, and edit only where you can edit.
Every client connects to the same server URL:
https://mcp.datacnvs.com/mcpDataCNVS is listed in the Claude connectors directory, so you can connect it from inside Claude with no setup. We're submitting it to the ChatGPT directory too; until it's listed there, use the generic settings below.
Claude (web, desktop and mobile) #
DataCNVS is in the Claude connectors directory, so connecting takes a few clicks. A connector you add on claude.ai is also available in the Claude desktop and mobile apps, under the same account. You don't need a paid Claude subscription to use the connector.
- Open the DataCNVS page in the Claude directory, or go to Settings → Connectors in Claude and search for DataCNVS.
- Choose Connect to Claude.
- Sign in with your DataCNVS account and approve the access Claude asks for: reading your canvases and editing them.
- Set tool permissions as recommended below.
- In a chat, turn DataCNVS on from the tools menu and ask away.
Recommended tool permissions #
By default Claude asks before each tool that changes something. Open Settings → Connectors → DataCNVS and set tool permissions. We recommend Always allow for all read and write tools, except those whose names start with the word delete. Leave those on Needs approval to prevent accidental deletions.
delete_canvasdelete_canvas_themedelete_observationdelete_observation_memberdelete_plot_objectdelete_variableRead and write tools are safe to automate: edits apply as you, show up live for collaborators, and can be undone for 7 days. Delete tools are final, so you stay in the loop for them.
On Claude Team and Enterprise, an Owner adds the connector once for the organization. Each member then connects it with their own DataCNVS sign-in, so everyone works with their own permissions.
Claude Code #
Add the server from your terminal:
claude mcp add --transport http datacnvs https://mcp.datacnvs.com/mcp \
--client-id 3f7or4t9so6894qi3ofgbgs79j --callback-port 33418Then start Claude Code, run /mcp, pick datacnvs and choose Authenticate. Your browser opens the DataCNVS sign-in, and Claude Code stores the token for you.
Keep --callback-port 33418 exactly as shown: DataCNVS only accepts sign-ins that come back to that port. If something else is using it, close that program and try again.
The command above adds DataCNVS to the current project only. To have it in every project, add --scope user:
claude mcp add --transport http --scope user datacnvs https://mcp.datacnvs.com/mcp \
--client-id 3f7or4t9so6894qi3ofgbgs79j --callback-port 33418Because Claude Code runs on your own computer, it also gets the request_upload tool, which uploads a file from your machine straight into the DataCNVS upload pipeline. The Claude web, desktop and mobile apps can't send files this way; upload in DataCNVS instead.
Generic MCP settings #
To connect any other MCP client, or to check a client's configuration, these are the server's settings:
| Setting | Value |
|---|---|
| Server URL | https://mcp.datacnvs.com/mcp |
| Transport | Streamable HTTP |
| Authentication | OAuth 2.1, authorization code + PKCE (S256) |
| Scopes | https://mcp.datacnvs.com/mcp/canvas.readhttps://mcp.datacnvs.com/mcp/canvas.write |
| Protected resource metadata | https://mcp.datacnvs.com/.well-known/oauth-protected-resource |
| Authorization server metadata | https://mcp.datacnvs.com/.well-known/oauth-authorization-server |
| Rate limit | 60 tool calls per user per minute |
| Undo window | 7 days after an edit |
The server doesn't support dynamic client registration, so a client needs a pre-registered OAuth client ID. Claude connects through the directory listing and Claude Code as described above. To connect something else, email support@datacnvs.com with the client's name and its OAuth redirect URL.
The read scope covers finding and reading canvases, observations, variables, themes and uploads. The write scope is needed for every edit and action. A client signed in with only the read scope is offered only the read tools.
To try the server by hand, MCP Inspector (npx @modelcontextprotocol/inspector@latest) works with the same client ID as Claude Code: choose Streamable HTTP, enter the server URL and the client ID, and sign in.
How the tools work #
A few things are worth knowing whichever client you use:
- Workspaces first. The AI starts by listing your workspaces and canvases, then reads a canvas's structure before editing it. You can name a canvas by title or paste its link.
- Canvas links work everywhere. Any tool that asks for a canvas accepts its DataCNVS URL.
- Edits apply immediately, as you. They appear live for anyone with the canvas open, like any other edit.
- Edits can be undone. Every canvas edit returns an edit ID, and the AI can undo it for 7 days. You can also restore an earlier version from the canvas's version history in DataCNVS.
- Some actions can't be undone. Deleting a canvas, observation or theme, revoking a share, and upload pipeline steps are final. Clients mark these tools as destructive and ask you first.
- You stay in control. Claude asks for your approval before tools that change things unless you've allowed them (see the recommended tool permissions); Claude Code follows your permission settings.
- Results link back. Results that name a canvas include its link, so you can jump straight to it.
Privacy: AI never sees your data values #
Scientific data is often unpublished, confidential or regulated, so DataCNVS is built to let an AI help you construct the figure without ever giving it the data in it.
AI clients build the plot. DataCNVS draws your data into it. The values themselves never leave DataCNVS.
What the AI cannot see #
- Data values. Series values (every point of every curve) and scalar values (single measurements, such as a sample's mass or a furnace setpoint) are never sent to an AI client. The tool that reads an observation returns no values, and every other tool's result has values removed before it leaves DataCNVS.
- Your uploaded files. No tool returns the contents of a file you uploaded.
- Anything you can't open. The AI works only with what your own account can access, in workspaces where MCP is turned on.
What the AI works with #
- The canvas structure: panels, axes and their ranges, series, legends, annotations, themes and styling.
- Names and descriptions: canvas and observation titles, variable names, units, quantity kinds and the text labels on your figures.
- References: IDs for observations and variables, which it passes back to DataCNVS to say which data goes where.
That's enough to build and restyle a figure, but it does mean titles, names and labels are visible to the AI. Keep anything sensitive out of them if that matters to you.
How figures still show your data #
When the AI plots a series, it hands DataCNVS a reference to the observation and variable, never the numbers. DataCNVS then loads the values in your browser, converts their units and draws them. Axis ranges, ticks and fits are all worked out by DataCNVS.
Text that shows a value, such as a label reading a sample's mass, uses a label token: a placeholder naming the variable, which DataCNVS fills in and converts to the right units when it draws the label. The AI writes the placeholder; it never sees what it becomes.
You can still ask for data by value, such as "observations where the anneal temperature was above 600 °C". DataCNVS runs that comparison itself and gives the AI only the IDs of the observations that match.
Good to know #
- Your AI provider sees your conversation. What you type, and the names and structure the tools return, are processed by the AI provider you chose, such as Anthropic, under your agreement with them.
- Local agents can read local files. Claude Code runs on your computer and can read files on it, independently of DataCNVS. If you upload a file with
request_upload, the agent has already got that file from your disk; DataCNVS still never returns its values. - Values you type are yours to share. If you paste numbers into the chat, the AI sees them, as it would anywhere else.
For more, see our privacy policy and security overview.
Usage examples #
Talk to the AI as you would to a colleague who knows DataCNVS. Each example shows a prompt, what happens, and the tools typically involved.
Find and describe a canvas #
What's on my "Creep tests 2026" canvas? List each panel and what it plots.
The AI finds the canvas in your workspaces and describes its layout: panels, axes, units and the series on each, by name. It describes the figure, not the numbers in it.
list_workspaceslist_canvasesread_canvas_modelBuild a figure from your observations #
Find my tensile tests on sample batch B-17 and make a new canvas plotting stress against strain, one line per specimen.
The AI searches your observations, creates a canvas and plots each one as a series by reference. DataCNVS loads and draws the data.
search_observationscreate_canvasplot_seriesChange units and axis titles #
On "Furnace run 14", show temperature in °C instead of K and title the y axis "Temperature (°C)".
Unit conversion happens in DataCNVS. The AI only chooses the display unit, so the converted values never pass through it.
read_canvas_modelset_axis_display_unitset_axis_titleAnnotate a plot #
Add a dashed threshold line at the yield stress on the top panel and a legend in the upper left.
Annotations, threshold lines, callouts and legends are all plot objects the AI can place and style. Values you give it in the prompt, like a threshold, are yours to share.
read_canvas_modelinsert_threshold_lineinsert_legendLay out a multi-panel figure #
Make this a two-by-two figure: copy the top-left panel into the other three.
The AI works with the canvas grid: inserting and duplicating panels, and setting gutters and margins.
insert_panelduplicate_panelconfigure_canvas_gridRestyle for a journal #
Apply our "Journal single column" theme to every canvas with "Nature Paper" in the title.
The AI lists your workspace themes and applies one to each canvas, so fonts, sizes and colours match across a paper.
list_canvas_themesapply_canvas_themeUpload a file (Claude Code only) #
Upload ./data/run_42.csv to DataCNVS, map its columns to our usual variables and tell me when it's ready to plot.
Claude Code uploads the file from your computer, waits for parsing and submits the unit mappings. DataCNVS does the parsing and conversion; the parsed values are never returned to the AI.
request_uploadwait_for_uploadsubmit_unit_mappingsUndo a change #
Undo the last two changes you made to that canvas.
Each canvas edit has an edit ID the AI can undo, as long as nothing has changed the same objects since. For bigger rollbacks, restore an earlier version from the canvas's version history.
undo_canvas_editTurning access off #
- For yourself: remove or disconnect DataCNVS in your client (Claude: Settings → Connectors; Claude Code:
claude mcp remove datacnvs). This deletes the client's tokens. - For a whole workspace: a workspace admin turns off MCP / external AI clients in Settings → Workspace → General → AI. It takes effect on the next tool call, for every member and every client.
- For shared data: the switch belongs to the workspace that owns a canvas or observation, so turning it off also covers that data wherever it has been shared.
Need help? #
If a client won't connect, a tool refuses, or you'd like client credentials for something not listed here, email support@datacnvs.com with the client you're using and what you asked it to do.