maidr-mcp

An MCP server that shows accessible maidr charts inside ChatGPT and Claude conversations, and lets the conversation's model move the reader through them and run the chart's commands for them.

The model calls show_chart with the data, and the chart appears in the conversation as an MCP App. A blind or low-vision reader Tabs into it and explores it the way they explore any maidr chart: arrow keys, screen reader, sonification, braille. Three things then happen through the model:

• The model can take the reader somewhere. When the reader asks for "the highest bar", the model calls maidr_get_layer_data to find it and maidr_navigate to move there, and maidr announces the point by speech, braille and sound. When the reader is typing to the model rather than in the chart, it passes focus: true, since they asked to be taken there, and maidr moves their keyboard focus into the chart to make the move.
• The model can press the chart's keys for the reader. When the reader asks to turn braille off, play the chart, or jump to the lowest value, the model calls maidr_list_commands to find the command and the reader's current modes, and maidr_run_command to run it. maidr announces the result as if the reader had pressed the key, and the model tells them which key that was.
• The model knows where the reader is. As the reader moves, the chart tells the model their position, so "what is this point?" needs no tool call. Within a turn, maidr_list_charts reads it live.

Experimental. Checked end to end in the MCP Apps SDK's reference host (below). Not yet checked in claude.ai or ChatGPT.

How it works

1. show_chart draws the chart on the server with matplotlib and py-maidr, so the model sends numbers rather than an SVG. The SVG reaches the chart view in the tool result's _meta, which is not added to the model's context.

2. The view loads maidr.js and the MCP Apps SDK from cdn.jsdelivr.net, the only domain its CSP declares, at pinned versions.

3. maidr.js registers its own WebMCP tools on document.modelContext. No host hands a view's tools to the model yet (ext-apps#797), so the view supplies document.modelContext itself. Each server tool of the same name relays its call to the view that is open:

   • the model calls the server tool;
   • the view polls an app-only tool, picks up the call, runs maidr's tool, and posts maidr's answer back;
   • that answer becomes the model's tool result.

   The poll is a long poll: the server holds it for up to 20 seconds until a call comes, so the call reaches the chart at once. If the host cuts or refuses a request held that long, the view falls back to short polls, which the server answers at once, one every 2 seconds, so the model's call is still answered within the 10 seconds the server waits for the chart. The view tries a long poll again after a minute, and waits twice as long after each one the host cuts.

   If hosts adopt WebMCP for MCP Apps, as ext-apps#798 proposes, maidr's tools reach the model directly and the relay can go.

4. On each key the reader presses in the chart, the view sends their position with ui/update-model-context.

5. update_chart draws a new chart for a viewId and relays it to that view, which swaps it in place: maidr lets go of the old chart, with any move or command still waiting for the reader in it, and binds the new one. When something was waiting, the chart's status line says it went, and update_chart tells the model so it can take back what it promised the reader. The server keeps the latest chart, so a view that missed the swap, out of view or mounted again by the host, catches up on its next poll.

6. The view never moves the reader's focus itself, except to put a reader back into the chart update_chart replaced. maidr moves it into the chart only when the model passes focus: true, which the server's instructions and the tools' descriptions tell it to do only when the reader asked to be taken somewhere or for something now, and never because they left the chart. Any other move or command the model makes while the reader is outside the chart waits for them, as does one with focus: true that could not take them in (focused: false). Until they enter the chart, the view's status line says so: "The assistant has a move waiting for you: Tab into the chart to hear it." After update_chart the same line says "Chart updated: ", until the reader enters the chart, by Tab or with focus: true; when a move or command waits as well, it says both.

Tools

show_chart and update_chart take one of these chart types, the ones py-maidr marks stable. Each maps onto the maidr layer type shown:

Every type also takes title, x_label and y_label. A pie has no axes, so there they name what the slices are and what the values measure.

The model reads every layer's points with maidr_get_layer_data. Where the last column says no, maidr gives the points no target: the reader moves through them with the keys, and maidr_navigate answers layer not navigable.

Example prompts

Once the server is connected, a reader asks in plain words; nobody names a tool. The prompts below are grouped by what the reader is doing, and each says which tools the model is expected to call. Paste the data into the conversation, or let the model use numbers it already has from earlier in the chat.

Show a chart

The model calls show_chart once, and the chart appears with the viewId the other tools use.

Explore the chart through the model

Have the model press the chart's keys

The model first calls maidr_list_commands to see what the reader can run and which modes are on, then maidr_run_command, and tells the reader which key does the same thing next time.

Change the chart in place

A reader who was in the chart stays in it, on the new chart, and hears that it changed.

Whole conversations

• Homework help. "I'm blind and studying for a statistics exam. Show a histogram of these 50 test scores, play it from left to right so I can hear the shape, then tell me whether it is skewed." The model shows the histogram, runs autoplay_forward, and answers from the layer data.
• A quick look at a dataset. "Here is a CSV of house prices and floor area. Show me a scatter with a trend line, take me to the most expensive house, and tell me how far it sits from the line."
• Following a stock. "Here are a stock's open, high, low and close for the last ten trading days. Show them as a candlestick chart. Which day had the largest range between high and low?" Then: "Show the closing prices as a line instead, and take me to that day."
• Checking an experiment. "Show box plots of the three groups' reaction times. Which group has the widest spread? Turn braille on and take me to that box." The model answers from the data and runs toggle_braille; a box chart cannot be navigated by the model, so it tells the reader the keys to reach that box.
• Reviewing a budget. "Show my spending as a pie chart, then as a bar chart sorted from largest to smallest, and take me to the largest category."

Use it

maidr-mcp is published three ways, at the same version:

• PyPI: maidr-mcp, run with uv. uvx maidr-mcp serves HTTP at 127.0.0.1:8000/mcp; uvx maidr-mcp --stdio serves a host that starts the server itself. uvx maidr-mcp@latest takes the newest release rather than one uv has cached.
• Container image: ghcr.io/xability/maidr-mcp, tagged with each version and latest. It serves HTTP at /mcp on port 8000.
• MCP Registry: io.github.xability/maidr-mcp, for clients that add servers from the registry.

What the server needs depends on how the host reaches it:

With a public address

# Make a token once, and keep it: it is the <token> in each host's URL below.
export MAIDR_MCP_TOKEN="$(python3 -c 'import secrets; print(secrets.token_urlsafe(32))')"
uvx maidr-mcp --host 0.0.0.0 --port 8000
# or
docker run -p 8000:8000 -e MAIDR_MCP_TOKEN ghcr.io/xability/maidr-mcp

The endpoint is /mcp, over Streamable HTTP, and with a token also /mcp/<token>.

• Run one instance. The relay keeps each chart's queue in memory.
• Restarts are tolerated. An open chart registers itself again on its next poll after a restart.

Claude. Add a custom connector: Customize > Connectors > Add custom connector, with https://<your-host>/mcp/<token>.

• Plans: custom connectors work on every plan; Free allows one.
• Team and Enterprise: an Owner adds it.
• Where charts render: on web, Desktop, and iOS/Android once the connector is added.

ChatGPT. Turn on developer mode, then create an app for https://<your-host>/mcp/<token>. Developer mode is available on the web for Plus, Pro, Business, Enterprise and Education accounts.

An access token

Without a token, anyone who can reach the server can draw charts. With one, every HTTP request but OPTIONS (CORS preflights, probes) has to carry it, and any other gets 401. Set it with MAIDR_MCP_TOKEN, with --token-file or MAIDR_MCP_TOKEN_FILE naming a file that holds just the token, or with --token. Over STDIO there is none: the host starts the server itself.

A request carries the token one of two ways:

• In the URL, as /mcp/<token>. Claude's custom connectors and ChatGPT's developer-mode apps take a URL and nothing else short of OAuth, so this is the form for them.
• In a header, as Authorization: Bearer <token> on /mcp, for clients that can send one.

python3 -c 'import secrets; print(secrets.token_urlsafe(32))' makes a good one. The server refuses a token shorter than 16 characters, or one holding anything but letters, digits and -._~.

• The URL is the secret. It lives in the host's connector or app settings, and anyone who sees it there can use the server.
• Rotate it by restarting the server with a new token, then give each host the new URL. The old one stops working at once.
• It is one shared token, not OAuth. Everyone given the URL shares it, and it cannot be taken back from one of them alone. Full OAuth is not implemented.
• Logs. The server's access log shows the URL as /mcp/<redacted>. A proxy, load balancer or platform in front of the server may log the full URL; use the header where the client can send one.
• --token shows in the process list to other users of the machine. Prefer the variable, or a file: with Docker secrets, -e MAIDR_MCP_TOKEN_FILE=/run/secrets/<name>.
• Serve it over HTTPS. Over plain HTTP the token crosses the network in the clear.

ChatGPT without a public address

Your own machine, through Secure MCP Tunnel. ChatGPT calls an app's MCP server from OpenAI's side, so it cannot reach localhost directly. OpenAI's Secure MCP Tunnel connects it to a server that stays off the internet:

• Server: runs on your machine, either maidr-mcp (HTTP at 127.0.0.1:8000/mcp) or maidr-mcp --stdio.
• Tunnel client: OpenAI's open-source client runs next to the server and opens only outbound HTTPS to OpenAI. It needs a tunnel_id from the Platform's tunnel settings and a runtime API key.
• App: you add it in developer mode through the tunnel, as that guide describes.
• Uptime: the machine has to stay on while the chart is in use.

The ChatGPT desktop app's own MCP servers. The desktop app can start a local server itself. Open Settings > MCP servers > Add server, choose STDIO, and give it this command, which needs uv:

uvx maidr-mcp --stdio

The desktop app shares this configuration with the Codex CLI and IDE extension. There are two caveats:

• Whether the app draws an MCP App's UI for a server added this way is not documented.
• This route has not been tried with maidr-mcp, so it is not known whether the chart appears.

Alongside the maidr skill

maidr-skill teaches AI agents to make the charts they write accessible, in Python, R or JavaScript. Both can be installed at once. The server's instructions and the skill split the charts the same way, and both tell the model to make each chart one way only:

The maidr_* tools here reach only the charts show_chart opened. A page the skill makes runs maidr.js, which, in a browser that supports it, offers tools of the same names to an agent there, without a viewId; the instructions tell the model which is which. MCP servers added in the ChatGPT desktop app reach the Codex CLI and IDE extension too, where the skill is often installed, so the split matters there most.

Try it in the reference host

e2e/run.sh checks the whole loop without a ChatGPT or Claude account. It does three things:

1. Builds the reference host from ext-apps (examples/basic-host, at tag v2.0.3).
2. Starts this server.
3. Drives both with Playwright the way a model and a reader would.

cd e2e && npm install && npx playwright-core install chromium && cd ..
bash e2e/run.sh

With MAIDR_MCP_TOKEN set, it runs the server behind that token, and the host reaches it at /mcp/<token>.

It checks:

• the chart appears;
• the model's calls are answered by maidr inside the chart, maidr_list_commands included, and maidr_run_command offers exactly the commands maidr lists as runnable;
• a move and a command made while the reader is in the chat wait, the chart's status line says so, and when the reader Tabs in, the move is announced, the command takes effect, and the notice goes;
• with focus: true, maidr_navigate takes a reader typing in the host page into the chart: their focus lands in it, the point is announced, the status line clears as on a Tab in, and their arrow keys move on from there. Within 10 seconds maidr moves their focus nowhere again: a move and a command with focus: true answer focused: false, wait, and the status line names them. After that a command with focus: true takes them in, answers queued, and runs after what waited;
• the arrow keys announce, and the reader's position reaches the host as model context;
• a move and commands made while the reader is in the chart take effect at once, and maidr_list_charts gives the model the position a command took the reader to;
• update_chart changes the chart in its own view, with no view added: a reader outside it keeps their focus, a reader in it stays in it, both are told, and maidr reads only the new chart; what waited for the reader in the old chart goes with it, and both the model and the status line say so, and a move waiting in the new one joins the change in the status line;
• a view that missed an update catches up on its next poll, and keeps to long polls;
• there are no console errors or CSP violations;
• a step, violin, pie and candlestick chart and a scatter with a trend line each appear: maidr reads each as the layers in the table above, ArrowRight announces its first point, and a move the model asks for is announced, or refused by maidr where the table says so;
• when the host cuts long polls, the view falls back to short polls, still answers the model, update_chart included, and stops polling when the chart is closed.

MAIDR_JS_FILE=/path/to/maidr/dist/maidr.js bash e2e/run.sh runs the same checks against a local build of maidr.js in place of the pinned release.

Limits

• The position the chart reports reaches the model with the reader's next message. Hosts apply ui/update-model-context on the next turn, so within a turn the model context misses the model's own moves and commands, and the reader moving on while the model answers. maidr_list_charts reads the position live, and the server's instructions tell the model to call it when the reader asks about "this point" and the context may be stale. What remains: maidr knows the position only while the reader is in the chart, so while they are typing to the model the last report stands; and a model that skips the call answers from the context alone.
• A move or command waits unless the reader asked to be taken there. Without focus: true, a move or command made while the reader is typing to the model is kept, and happens when they Tab back into the chart: the move first, then the commands, half a second apart. The tool answers applied: "on-next-focus" and the model is told to say so; the chart's status line says so too, where a screen reader browsing the conversation finds it. With focus: true, which the model is told to pass only when the reader asked, maidr moves their focus into the chart and does it there. What remains:
  • Safari refuses. WebKit keeps a chart in another page's frame from taking focus without the reader's own key or click, so there maidr answers focused: false, and the reader Tabs in themselves and finds the move or command waiting. Chromium and Firefox let the chart take focus.
  • Once every 10 seconds. maidr moves focus for the model at most once every 10 seconds on a page, so that a reader who left the chart is not pulled straight back in. A call sooner answers focused: false and waits like one without focus. Each chart view is a page of its own, so the limit does not span charts: the model could take the reader into a second chart at once, and only its instructions stop it.
  • The host's own dialogs. The chart cannot see a dialog the host shows over the conversation, so maidr cannot keep from taking focus from one. The model is told not to pass focus: true while one is open; nothing else stops it.
  • At most 8 commands wait.
• Each show_chart call still adds a chart. Claude mounts a new view for every call to a tool with a UI, and keeps the earlier ones. A chart that changes stays in its view only when the model calls update_chart, as the server's instructions ask; a second show_chart is a second view.
• The relay polls. Each open chart makes a request every 20 seconds, or every 2 seconds on a host that cuts requests held open, where a model's call can also take up to 2 seconds longer to reach the chart. A host that also allows a view fewer than 30 calls a minute leaves some of the model's calls unanswered. Which hosts limit calls from a view is not yet known. The same chart open twice, say in two tabs, makes each copy read the other's polls as cuts, so both settle on short polls.
• Access is one shared token, and only if you set one. Without a token, anyone with the server's URL can draw charts. Set MAIDR_MCP_TOKEN and every request needs it; Claude and ChatGPT carry it in the URL. That URL, kept in the host's connector settings, is then the secret; rotating it means restarting the server with a new token and updating each host. Full OAuth is not implemented. See An access token. A chart can only be read or driven with its viewId, a random 24-character token.
• Ten chart families, and no plotting code. The model sends data for one of the types above, and the server draws it. It deliberately takes no plotting code: anyone with its URL could run code on it. py-maidr's experimental plot types are left out until they have been tried with readers. The maidr skill makes the rest.
• show_chart cannot tell whether its chart appears. Its answer reads the same in every host, since a stateless request does not always say what the client can show. In a host that shows no MCP Apps, such as a terminal, the answer still says the chart is showing. The instructions and the tool's description tell the model not to call it there; nothing else stops it.

Network and data

• The server: receives the data the model sends to draw a chart. It keeps the chart's latest SVG in memory until the chart has gone 15 minutes without polling, and logs nothing about the data beyond the HTTP access log, which shows a token in the URL as /mcp/<redacted>.
• The chart view: loads maidr.js and the MCP Apps SDK from cdn.jsdelivr.net. maidr's own AI chat inside the chart works as it does anywhere else, with a key the reader adds.

Development

uv sync
uv run ruff check . && uv run ruff format --check .
uv run pytest
bash e2e/run.sh

How maidr releases reach maidr-mcp

• maidr.js, in the chart view, is the version MAIDR_JS_VERSION in server.py names. update-maidr.yml raises it within about an hour of an npm release, once the checks below pass, so an install or image made from main after that loads it, and the next weekly release takes it to PyPI and GHCR.
• py-maidr, which draws the chart, is resolved afresh by every uvx or Docker install, within maidr>=1.26,<2, so a new install has a release as soon as PyPI does. uv.lock, which the same workflow raises, pins the py-maidr that CI and a checkout test against.

.github/workflows/update-maidr.yml checks npm and PyPI every hour. A new maidr.js must carry npm provenance from maidr's release workflow, with a Sigstore signature that npm audit signatures verifies. Then ruff, pytest on Python 3.10 and 3.13, and e2e/run.sh run against the new versions, and only when all of them pass does the workflow commit them to main: as fix(deps): when the maidr.js pin moves, since that ships to users, and as chore(deps): when only uv.lock does. A maidr.js that changes its runnable commands fails e2e/run.sh rather than shipping with a stale RunnableCommand. When a check fails, nothing is pushed, and an issue names the versions, the step, and the run. A failure that may pass next time, such as a download or the browser's install, is tried again every hour for a day; one that would only repeat, such as a check failing on the new versions, holds them until the issue is closed or the workflow is run by hand. The issue closes itself once an update at least as new reaches main. Every update pairs npm's latest maidr.js with PyPI's newest py-maidr, so while a maidr.js release fails, a new py-maidr is tried only with it and uv.lock waits too; the issue says how to raise py-maidr alone, by running the workflow by hand with the maidr.js main already loads.

It commits to main itself rather than opening a pull request, by design. maidr.js and py-maidr are reviewed and tested in their own repositories before they are released. What this repository has to check is that its server, its view and the reader's way through the chart still work with them. That means announcements, focus and braille in e2e/run.sh, and the workflow checks all of it before it commits. A pull request opened with the workflow's own token would start no workflow, ci.yml included. It would only wait for a person to merge what the same checks had already passed, and the maidr.js fix it carries would wait with it. To look at each update yourself, disable the workflow and raise the pins in pull requests with scripts/update_maidr.py.

To pick another version, run the workflow from the Actions tab with a maidr.js version, or in a checkout run uv run --no-project python scripts/update_maidr.py update --maidr-js <version> --allow-lower and open a pull request. The next hourly run raises a lower pin again, so to hold maidr.js at an older version, disable the workflow in the Actions tab until maidr releases a fix.

License

GPL-3.0-or-later, like the rest of maidr.