MCP App (Dialog Lab), light

MCP App (Dialog Lab), dark

Colour picker (pick_color), Wheel, dark

Values, light

Palettes, dark

Tool

MCP feature exercised

Spec version that introduced it

inspect_client

Reports the protocol version, client info and declared capabilities (elicitation form/url, MCP Apps UI, sampling, roots, tasks, extensions)

n/a (reads initialize or the per-request _meta envelope)

list_seen_clients

Every distinct client this server process has seen, with capability flags and request counts

n/a

elicit_confirm

Form elicitation: a yes/no confirmation dialog; accept, decline and cancel are each reported

Form elicitation: 2025-06-18

elicit_all_field_types

Form elicitation with every allowed field type: string with min/max length, email/uri/date/date-time formats, integer, number, boolean, untitled and titled single-select, legacy enumNames, multi-select (untitled and titled), and defaults. variant: all, basic, formats or enums

Base types: 2025-06-18. Titled/multi-select enums and defaults on all types: 2025-11-25

elicit_parallel

Two independent elicitations in a single round

Multiple inputRequests: 2026-07-28. On 2025-era connections the SDK sends them one after the other

elicit_wizard

Three chained dialogs. Step 2 depends on step 1, and progress travels in HMAC-signed requestState

input_required + requestState (multi round-trip): 2026-07-28. Emulated on 2025-era connections

elicit_url

URL-mode elicitation: the client opens a local page, the user enters a fake token, and the server confirms it received the token (masked)

URL mode: 2025-11-25. As an input_required request: 2026-07-28

elicit_url_legacy_push

URL mode in the 2025-11-25 push style: elicitation/create, then notifications/elicitation/complete once the page is submitted. Returns an error on 2026-07-28 connections by design

2025-11-25 only

slow_task

Progress notifications, log messages, and cancellation (steps, delayMs)

Progress and cancellation: 2024-11-05. Logging notifications are deprecated in 2026-07-28

open_dialog_lab

MCP App: tool _meta.ui.resourceUri, then a ui:// resource rendered in an iframe (colour picker, slider, star rating, date, drag-to-rank list, sketch canvas). Buttons exercise tools/call, ui/message, ui/update-model-context, ui/request-display-mode and ui/download-file

MCP Apps extension io.modelcontextprotocol/ui (ext-apps spec 2026-01-26)

dialog_lab_submit

App-only tool (_meta.ui.visibility: ["app"]). The UI calls it and hosts should hide it from the model

MCP Apps

dialog_lab_upload

App-only tool behind the Dialog Lab's File upload card: takes a base64 file, returns its size and SHA-256, stores nothing. Size cap UPLOAD_MAX_BYTES (10 MiB)

MCP Apps (app → server tools/call)

dialog_lab_results

Returns what was submitted in the Dialog Lab, with the sketch as an image

n/a

return_files

One tool result with every way to hand over a file: embedded text resource (CSV), embedded blob resources (PNG, PDF), a resource_link to test://files/linked-notes.md, an image block and an audio block (WAV). The first text block lists them, so you can see what the client dropped

resource_link and audio: 2025-06-18. Embedded resources and images: 2024-11-05

Resource template test://files/{name}

The sample files behind return_files (sample.csv, gradient.png, sample.pdf, beep.wav, linked-notes.md), listed in resources/list and resources/templates/list, readable with resources/read, names completable

Resource templates: 2024-11-05

elicit_format_probe

Form elicitation with schema features outside the spec, one dialog per probe: a pattern, format color, format time, a nested object, a free-text array, a file-like data-url field. Reports per probe whether the user accepted (and what came back), declined or cancelled. variant: all or one probe

Not in any spec version (deliberately)

upload_file_experimental

Experimental, not in the spec: a file argument marked with x-mcp-file per the draft files extension (io.modelcontextprotocol/files), inline base64 only. Returns size and SHA-256 and checks a supplied digest

Draft extension, not a SEP

pick_color

MCP App used like a question: a colour picker (ui://feature-test/color-picker.html; prompt, initial #rrggbb or #rrggbbaa) with tabs Wheel (hue ring + S/V triangle), Classic (S/V square), Values (RGB/HSL/HSV/alpha), Pick (EyeDropper API, system dialog, sampling from a local image) and Palettes, plus opacity, previous colour and Recent (structuredContent.recent from earlier answers, since the sandboxed iframe has no storage). Drag, touch and keyboard, HiDPI canvases, light/dark. The call returns at once with a requestId; the answer comes back in a later step through pick_color_submit (app-only, #rrggbb or #rrggbbaa), ui/update-model-context and ui/message, each reported in the UI. See server-vs-app.md

MCP Apps extension

pick_color_result

Returns the colour the user picked as hex, rgb()/rgba() and hsl()/hsla() (or "cancelled" / "still waiting") for a requestId, or for the latest pick_color

n/a

view_3d

MCP App with WebGL: an interactive three.js viewer (ui://feature-test/viewer-3d.html; shape knot/sphere/gear/glb-sample, optional camera {position, target} and highlight). Orbit, zoom, auto-rotate, wireframe, click to select a part; the selection goes back through view_3d_select (app-only), ui/update-model-context and a Tell the assistant ui/message. glb-sample is read through the host with resources/read. Its Details panel probes WebGL/WebGPU/WASM/eval/fetch/fullscreen in the host's iframe. See server-vs-app.md

MCP Apps extension (app → server resources/read, ui/request-display-mode)

view_3d_state

Returns the part the user selected in a view_3d view and the camera at that moment, or "nothing selected yet"

n/a

Resource test://models/sample.glb

A small glTF binary (model/gltf-binary) generated in code, with the named meshes base, tower, roof, flag

Resources: 2024-11-05

Prompt try_all_dialogs

A prompt that walks the assistant through every tool above

Prompts

Mode

Command

Endpoint

stdio (what desktop clients launch)

node dist/stdio.js or npm start

stdin/stdout. URL-flow pages are served on http://127.0.0.1:3458/flow/…

Streamable HTTP

node dist/http.js or npm run start:http

http://127.0.0.1:3457/mcp. URL-flow pages are served on the same port

Docker

see docs/docker.md

Variable

Default

Mode

Meaning

PORT

3457

HTTP

Listen port. 0 picks a free port, and the actual URL is logged.

HOST

127.0.0.1

HTTP

Bind address. Use 0.0.0.0 in containers.

PUBLIC_URL

http://<HOST>:<PORT>

HTTP

Externally visible base URL, without /mcp. Used for URL-elicitation links. Also enables localhost Host/Origin validation when it points at localhost/127.0.0.1.

ALLOWED_HOSTS

(unset)

HTTP

Comma-separated hostnames accepted in the Host and Origin headers, for example mcp.example.com. When unset, localhost names are enforced if HOST or PUBLIC_URL is local; otherwise the PUBLIC_URL hostname plus the localhost names are accepted. Set it when clients use another name (for example a Docker service name).

FLOW_PORT

3458

stdio

Port for the URL-elicitation pages. Falls back to a free port (with a warning) if it is busy. 0 always picks a free port.

FLOW_HOST

127.0.0.1

stdio

Bind address of that listener. Use 0.0.0.0 in Docker.

FLOW_PUBLIC_URL

http://<FLOW_HOST>:<port>

stdio

Base URL that the URL-elicitation links point to, when the browser has to use a different address or port.

FLOW_WAIT_MS

300000

both

How long elicit_url / elicit_url_legacy_push wait for the browser page after the client already answered accept (clients may accept when the link is opened).

UPLOAD_MAX_BYTES

10485760 (10 MiB)

both

Largest decoded file dialog_lab_upload and upload_file_experimental accept. Also sizes the HTTP request-body limit (this value as base64 plus 1 MiB; larger bodies get 413) and the stdio message buffer.

LOG_LEVEL

info

both

debug, info, warn, error or silent.

LOG_FORMAT

text

both

text (readable, multi-line capability tables) or json (one JSON object per line).

Tool

In a client with full support

If the client lacks the feature

inspect_client

A capability table (elicitation form/url, MCP Apps UI and its mimeTypes, sampling, roots, tasks) followed by the raw JSON

Always works

elicit_confirm

A dialog "Really delete …?" with one checkbox. The result says ACCEPTED, DECLINED or CANCELLED

Tool error: "… did not declare form-mode elicitation support …"

elicit_all_field_types

One form containing about 13 fields of every type, with defaults pre-filled. The result lists each value and its JSON type, which shows whether the client sends integers as numbers, arrays for multi-selects, and so on

Same tool error

elicit_parallel

Two dialogs ("what should I call you?" and "how are you?"), shown together or one after the other

Same tool error

elicit_wizard

Three dialogs in a row. The step 2 limit depends on the size chosen in step 1, and step 3 shows a summary

Same tool error

elicit_url

The client offers to open a URL. The page (below) asks for any token. Submit it, go back, accept, and the result reports a masked token

Tool error mentioning url-mode. If you accept before submitting the page, the tool waits for the page and returns an error after FLOW_WAIT_MS if it is never submitted

elicit_url_legacy_push

2025-era connections only: the same flow, and the client should close its dialog by itself when it receives notifications/elicitation/complete

On 2026-07-28: a tool error explaining that push-style elicitation does not exist there

slow_task

A progress indicator counting up. Cancelling stops it early

Result text: "Client did not request progress."

open_dialog_lab

The Dialog Lab UI (screenshots above)

Only the text result, which says that no UI appeared because the client does not render MCP Apps

dialog_lab_results

The last submissions as JSON, plus the sketch as an image

Dialog Lab File upload

Pick a file, press Upload to server: the app shows size and SHA-256 (and whether they match the local file). Try 100 KB, 1 MB and 10 MB files to find the host's message-size limit

An error under the upload card: the host's or the server's limit, or no serverTools support

Dialog Lab downloads

Three buttons: JSON (embedded text), sketch PNG (embedded blob) and linked PDF (resource_link, which the host has to read from the server)

All three are disabled, with a note, when the host doesn't advertise hostCapabilities.downloadFile

return_files

Seven blocks: the list, a CSV, a PNG, a PDF, a link to a Markdown file, an image and a short beep

Whatever is missing from the list was dropped (or handed only to the model)

elicit_format_probe

One dialog per probe. The result says per probe what came back; every one of these is outside the spec, so a fallback text box, a dropped field or a refusal is all legitimate

A client that refuses a schema ends the whole call with an error (the official TypeScript SDK client refuses everything except pattern). Use single variants to find which probe it refuses

upload_file_experimental

Only clients implementing the draft files extension render a file picker; others show the raw object

A model can still call it with inline base64

pick_color

A colour picker panel (screenshots at the top) with the question, OK and Cancel. After OK the assistant learns the colour from the model context and/or a chat message ("I picked #xxxxxx for: …", #xxxxxxxx with opacity); the picker lists which of the three channels worked

Only the text result; pick_color_result then keeps answering "still waiting"

view_3d

A 3D view (screenshots below) you can drag, zoom and click. Selecting a part reports Server: stored, model context and (after Tell the assistant) the chat message; view_3d_state then names the part. Fullscreen appears only if the host offers it

Only the text result ("If no 3D view appeared, this client does not render MCP Apps."). With glb-sample, a host without resources/read for apps shows an error inside the viewer

3D viewer (view_3d), light

3D viewer (glb-sample), dark

1. The picker, rendered inline under the tool call

2. OK: confirm the app's call to pick_color_submit

3. Confirm the chat message ("I picked …")

4. The model reads the answer with pick_color_result

The Pick tab: a loaded image, a sampled colour

Symptom

Check

No dialog appears

Run inspect_client, or look at the client.connected/client.seen log line. If elicitation form: no, the client does not declare elicitation, and the tools return an explanatory error instead of a dialog. Some clients show elicitation only in agent mode or after you approve the tool call.

elicit_url says the client doesn't declare url

The client only supports form mode (elicitation: {} means form only).

The URL page doesn't load

In stdio mode the page is served on FLOW_PORT (default 3458; check the flows.listening log line for the actual port). In Docker, publish that port and set FLOW_HOST=0.0.0.0. Behind a proxy, set PUBLIC_URL (HTTP mode) or FLOW_PUBLIC_URL (stdio mode).

elicit_url seems to hang after you accepted

Some clients (e.g. Open WebUI) answer accept as soon as the link is opened; the tool then waits for the browser page to be submitted. Submit the page, or the tool returns an error after FLOW_WAIT_MS (default 5 min).

No graphical UI for open_dialog_lab

Check MCP Apps UI in inspect_client. The client must declare the io.modelcontextprotocol/ui extension with text/html;profile=mcp-app.

HTTP answers 403

Host/Origin validation (DNS-rebinding protection). Add your hostname to ALLOWED_HOSTS.

HTTP answers 413, or a Dialog Lab upload fails

The request body was above UPLOAD_MAX_BYTES as base64 plus 1 MiB, or the file was above UPLOAD_MAX_BYTES. Raise the variable. If the server never logs upload.received/upload.rejected, the host refused the message before it reached the server.

The URL page answers 409

The flow was already submitted once. Run the tool again for a new link.

HTTP answers 400 "Missing session" or 404 "Unknown session"

A 2025-era client sent a request without initialize first, or after the server restarted. Reconnect.

elicit_url_legacy_push fails

Expected on 2026-07-28 connections. Use elicit_url.

The client shows garbage or JSON parse errors (stdio)

Something wrote to stdout. This server never does; check wrapper scripts.

dist/ui/dialog-lab.html missing ("Dialog Lab UI not built")

Run npm run build.