Show widget

show_widget is a core tool that shows a self-contained HTML widget on the user's current surface. OpenClaw renders it inline in the Control UI and in iOS, Android, macOS, and Linux Quick Chat transcripts. The Linux dashboard uses the browser Control UI. In a Discord session with Activities enabled, the Discord plugin posts an Open widget button that launches it as an Activity.

For a pinned data report, provide a structured report with pin: true. Reports render directly in the Control UI dashboard using its native typography and layout, without a document frame. Use HTML for arbitrary interactive content or an inline preview. See Native dashboard reports.

How widgets work

For HTML widgets, OpenClaw core validates widget_code by parsing every inline JavaScript <script> (classic and module), skipping scripts with src or a non-JavaScript type, then wraps it once in the canonical HTML document. Core rejects the call with the line and column of the first syntax error, so a widget with a broken script is never hosted. For an inline client, core stores that document as a Canvas document and returns a preview handle. The Control UI reads the document over its authenticated Gateway connection and renders it through the dedicated-origin, double-iframe sandbox used by dashboard widgets and MCP Apps. The widget frame does not need its own login session. iOS, Android, macOS, and Linux Quick Chat use isolated web views. Full chat clients restore the widget after history reload. Quick Chat keeps the widget for its active reply.

Channel plugins can register a contextual presenter behind the same core tool. In a configured Discord session, core hands the composed document to the Discord presenter, which stores it and posts the Activity button in the current channel. The model still makes one show_widget call. There is no transport-specific widget tool or content kind.

In Control UI sessions, a Canvas widget can also be pinned to the session dashboard. Set pin: true in the tool call, or use Pin to dashboard on an existing transcript widget. Pinning gives the dashboard copy its own identity and capability grants. The inline preview never inherits those grants. The browser never resolves a widget data binding inside the untrusted frame.

For browser embedding, the wrapper document injects six small host bridges around the widget code:

• A size reporter posts the rendered content height to the embedding chat. The chat clamps that height and fits the iframe (48 to 8000 pixels).
• A host bridge defines a global sendPrompt(text) helper plus the structured openclaw.prompt, openclaw.state, openclaw.data, and openclaw.cron APIs. sendPrompt(text) is the fire-and-forget form of openclaw.prompt.send. Inline chat prompts retain their private message channel. Dashboard APIs use a view-ticket-bound request channel. See Interactive widgets and Dashboard capabilities.
• An error reporter captures uncaught script and event-handler exceptions and unhandled promise rejections. It sends at most three distinct messages per document load, with messages capped at 500 UTF-16 units, source basenames at 200, and optional integer line and column numbers. The Control UI shows a notice and forwards one report per document and chat session per page load to the Gateway as a session wake event, with an additional shared limit of 10 reports per key per 60 seconds (up to 100 tracked keys). Reports are forwarded to the agent only for widgets rendered within ten minutes of their message. Older restored history shows the notice without waking the agent. If the session already has an active run, the event stays queued and the immediate wake retries until the session lane is free, so the model sees it on its next available turn. The wake turn runs without the originating client capabilities, so show_widget can be unavailable there. The report asks the model to reply with the corrected code and show it on the next turn. Native apps do not report runtime errors yet.
• A theme bridge listens for the Control UI's current design tokens and applies them as CSS variables. It does this on load and again on every theme change.
• A snapshot bridge renders the current widget document as a PNG when the embedding chat requests an export.
• A chat-host bridge hides embedded scrollbar chrome when the widget runs inline while preserving scrolling behavior.

Everything else stays inside the frame. The document runs in an opaque origin with a strict Content Security Policy. Widget scripts cannot reach the Control UI, the Gateway, or the network.

OpenClaw exposes show_widget only when the originating Gateway client declares the inline-widgets capability or exactly one registered current-channel presenter synchronously matches trusted run context. The Control UI and supported native apps declare the inline capability automatically. Linux Quick Chat stays text-only for Gateway connections that require a custom TLS leaf pin because its platform WebView cannot bind that pin. Discord matches only when Activities are configured for the current account and a concrete channel is available. Other channel runs without an inline client or matching presenter do not receive the tool.

An agent-turn automation bound to a persistent session and carrying a server-authored scheduled tool policy may explicitly allow show_widget without an inline client. That scheduled surface is pinned-only: every call requires pin: true, writes to the bound session dashboard, and cannot set presentation.target. Detached cron-run sessions, ordinary capless channel runs, and scheduled jobs without an explicit tool cap remain excluded. The originating-client capability remains mandatory for inline presentation.

When the Gateway automatically resumes an interrupted Control UI turn after a restart, the recovered turn can also create or update pinned dashboard widgets without a connected browser. Recovery uses the same pinned-only surface: set pin: true and omit presentation.target. Inline previews still require a new turn from a client that declares inline-widgets. The resumed turn does not inherit a browser connection or device presentation rights.

Capability transport covers embedded, Codex app-server, and CLI-backed model backends. Grant-authenticated MCP callers without inline-widgets remain fail closed unless their trusted run context matches a presenter. Authenticated direct HTTP tools/invoke requests cannot request inline rendering, but a request carrying eligible current-channel context can use the matching presenter. Authentication never bypasses presenter or route eligibility.

Design system

Every Canvas widget includes a classless base stylesheet and a small token set:

Bare headings, paragraphs, links, buttons, inputs, selects, textareas, tables, and code blocks receive base styles. Helper classes provide common patterns:

• .card for a bordered content surface
• .badge, plus .ok, .warn, .danger, or .info, for compact status labels
• .metric for a prominent numeric value
• .muted for secondary text
• .row for a wrapping horizontal layout
• button.primary for the primary action

The Control UI posts an openclaw:widget-theme message with the active theme values when a widget loads. It posts again whenever the theme changes. Widgets therefore track every theme family, including Claw, Knot, Dash, and custom themes, without reloading. Outside the Control UI, including native apps and direct opens, widgets use the baked light or dark palette selected by prefers-color-scheme.

Author widgets with four rules:

1. Use the design variables for every color and background. Do not hardcode color values.
2. Keep the page background transparent so the widget belongs to its host surface.
3. Reserve --accent-fill for at most one primary action.
4. Fit the iframe width at every viewport. Avoid fixed page or card widths. Use fluid sizing and wrap or stack multi-column layouts when narrow. Use horizontal scrolling only when exact geometry must remain.

Export: In web chat, open the widget card menu. From there, copy the rendered widget to the clipboard or download it as a PNG. Older widget documents without the snapshot bridge fall back to an HTML file download.

Use the tool

The core tool requires title and one content input: widget_code for HTML or registered source, or report for a native dashboard report.

Discord also accepts optional button_label text for the Activity launch button. The Canvas schema intentionally omits this Discord-only field.

The core show_widget tool also accepts these optional dashboard placement fields, including when Discord is the presentation destination:

• kind: html by default, plus active registered source kinds when available. Applies to widget_code. Omit with report.
• pin: also place an HTML or registered widget on the session dashboard. Required for reports and on the pinned-only scheduled surface.
• name: stable widget name. Defaults to a slug of title.
• tab: destination tab slug.
• size: one of sm, md, lg, xl, or full.
• presentation.frame: pinned dashboard frame: card, full-bleed, or frameless.
• after: sibling widget name after which to place the widget.
• capabilities: access requested by a pinned widget. netOrigins contains exact HTTPS origins. tools contains prompt, an allowlisted read binding, or an exact cron.trigger:<jobId> action.

An inline result includes a Canvas preview handle, so the Control UI and supported native apps render the widget directly from the tool call and restore it after history reload. A successful current-channel presentation returns a generic message receipt describing what became visible. Pinned results retain the board widget name so the Control UI does not offer a duplicate pin after transcript reload.

Pinned results also report capabilityState: none, pending, rejected, or granted. A saved widget is not necessarily authorized to read data. For pending, ask the operator to review the dashboard permission card. For rejected, review the requested access and session permission policy with the operator before retrying. The inline preview does not inherit dashboard grants.

If current-channel presentation fails, core falls back inline only when the originating client actually supports inline widgets. Otherwise the tool fails visibly. When pin: true succeeded before presentation failed, the result is explicitly partial and names the durable board widget. Presentation failure never rolls back that unrelated board state.

Native dashboard reports

Use a report for a pinned summary made of text, metrics, tables, simple charts, and links. Provide the report object directly in report:

{  "title": "Weekly delivery report",  "name": "weekly-delivery",  "pin": true,  "size": "lg",  "report": {    "blocks": [      {        "type": "metrics",        "items": [{ "label": "Completed", "value": "24", "detail": "This week" }]      },      {        "type": "table",        "columns": ["Team", "Completed"],        "rows": [          ["Platform", "14"],          ["Product", "10"]        ]      }    ]  }}

The result names the pinned board widget. It creates no inline preview, native device panel, or Discord Activity. Do not combine report with widget_code, kind, capabilities, or device presentation. The report field does not change the registered source-kind namespace.

Each block has a required type:

A report contains 1–24 blocks and at most 8KB of encoded JSON. Titles have 1–120 characters, labels 1–160, details at most 240, and URLs at most 2,048. Chart values stay between -Number.MAX_SAFE_INTEGER and Number.MAX_SAFE_INTEGER. Unknown fields are rejected. Reports accept no HTML, CSS, scripts, media, executable actions, network reads, or Gateway RPCs. Links are ordinary navigation.

Reuse the same name and pin: true with a new report object to replace a report's data. The dashboard tool can also create or update it with action: "widget_put", pluginKind: "session:report", and props: report. To convert an existing HTML widget, first remove it with dashboard action widget_remove, then create the report. A same-name update cannot change the widget's content owner.

Show on a device

When a widget presenter plugin is active, presentation.target also offers node_panel. OpenClaw creates the same hosted widget document, selects a connected widget-panel-capable Mac, and opens its native panel at that document. The tool result names the selected Mac.

If no eligible Mac is connected or the node command fails, the widget still appears inline in chat and the result explains how to recover. Pair a Mac running OpenClaw or open the macOS app, then retry. Widgets shown in a native panel are render-only. Widget actions are disabled there.

Interactive widgets

In the Control UI, widget scripts can drive the conversation. The wrapper document defines a global sendPrompt(text) function. Calling it submits text to the chat as if the user had typed and sent the message. Wire it to buttons or other controls to build interactive flows such as pickers, quizzes, or drill-down dashboards. Native apps render interactive widget code but do not expose this chat prompt bridge.

<button onclick="sendPrompt('Show the failing tests in detail')">Failing tests</button>

Every prompt is validated on both sides of the frame boundary:

• sendPrompt requires transient user activation inside the widget: it only works in the few seconds after the user clicks or presses a key in the widget, so wire it to buttons and other click targets — calling it automatically on load does nothing. The bridge keeps the sending endpoint private to itself and fails closed in browsers that do not expose user activation, so widget code cannot bypass the check.
• Prompt authority belongs to the original widget document only. The trusted bridge offers its channel endpoint to the chat before widget code can run or navigate the frame, the chat adopts only that first offer, and the channel dies with the document on navigation. Externally allowed embed URLs are never adopted.
• The widget frame must be visible in the chat transcript and hold focus. That is an additional host-observed signal that the user is actually interacting with this widget.
• The text must be non-empty after trimming and at most 4,000 characters.
• Prompts starting with / are rejected, so widget code cannot trigger chat commands such as /approve or /stop.
• Each widget document may send at most 10 prompts per rolling minute. Excess prompts are dropped silently.

Accepted prompts appear in the transcript as regular user messages and start a normal agent turn in the session that owns the widget. There is no feedback channel into the widget: a dropped prompt fails silently, and the widget cannot read the agent's reply.

Dashboard capabilities

Pinned HTML and registered-source widgets expose one ticket-bound host API. An explicit session permission mode determines how declared capabilities are approved: Full access grants them immediately. Workspace uses an AI reviewer and rejects anything it does not allow. Guarded shows an Allow / Reject card. Read only rejects them. Without an explicit session mode, the equivalent configured exec approval policy applies.

• openclaw.host.controlUiBaseUrl exposes the Control UI origin plus its configured base path after the dashboard host initializes. It is null before initialization and outside the dashboard, so read it in the link's click handler rather than when the widget script first runs.
• openclaw.prompt.send(text) requires transient user activation and posts a visible composer message. Declaring and receiving the prompt tool grant skips the extra per-click confirmation. Validation, focus checks, and rate limits still apply.
• openclaw.state.emit(payload) adds a session notice. Payloads are capped at 8 KiB, and identical client emissions within five seconds are coalesced.
• openclaw.data.read(bindingId, params?) resolves only at the Gateway. Core bindings are sessions.list, usage.status, usage.cost, cron.list, cron.status, agents.list, health, and the repository-scoped github.actions.runs binding below.
• openclaw.action.run(actionId, params?) invokes an operator-granted plugin dashboard action verb through its write-scoped Gateway method.
• openclaw.cron.trigger(jobId) runs an existing job now only when the exact cron.trigger:<jobId> capability was granted.

OpenClaw forwards user-clicked links to http or https destinations to the Control UI host. The host opens a new tab with noopener and noreferrer. Forwarding covers a primary click on a target="_blank" link and a middle-button click on any link, matching how links behave elsewhere in the Control UI. A widget's own preventDefault still cancels the click. The widget sandbox never grants popup permission, and script-initiated window.open does not work.

Network access is separate from host tools. Put exact HTTPS origins in capabilities.netOrigins. Once the session policy grants them, only those origins enter the widget's connect-src. Wildcards, credentials, paths, query strings, and undeclared origins remain blocked. A literal port is allowed only when it is part of the declared origin.

The tool schema describes currently active plugin read bindings and action verbs, including action parameter schemas when provided. This discovery text is bounded: it includes complete entries and reports when entries were omitted. Disabled plugins are not advertised.

Read GitHub Actions runs

With a usable connected agent GitHub identity, use the host binding instead of fetching GitHub directly from widget code. For example, give show_widget this input, replacing owner/repo in both places:

{  "title": "Workflow runs",  "name": "workflow-runs",  "pin": true,  "capabilities": { "tools": ["github.actions.runs:owner/repo"] },  "widget_code": "<pre id='runs'>Loading…</pre><script>openclaw.data.read('github.actions.runs',{repository:'owner/repo',perPage:20}).then(data=>{document.getElementById('runs').textContent=data.workflow_runs.map(run=>run.display_title+': '+(run.conclusion||run.status)).join('\\n')}).catch(error=>{document.getElementById('runs').textContent=String(error)});</script>"}

The read runs in the pinned dashboard after approval, not the inline preview. No netOrigins declaration or token is needed. Repository names are case-insensitive and grants use lowercase spelling. A grant for one repository cannot read another. Granting only https://api.github.com network access does not authorize this binding.

Direct browser fetches to api.github.com do not inherit the agent's login and share GitHub's anonymous IP quota. Existing widgets using fetch should replace that call with the binding and replace their GitHub netOrigins declaration with the repository-scoped tool grant. Check the effective account in Agent settings → Tools → GitHub account. Administrators can also reach it from Settings → Profile → GitHub connections.

For refreshable widgets, keep the last successful data and a separate error element. Clear the error after a successful read, and keep result containers in the document so later refreshes can recover. Show a Refresh control for an immediate retry.

Approval shares Actions metadata with the widget and its session audience, including metadata from private repositories accessible to the selected agent identity. The host selects the agent override, then the configured System identity, then native GitHub authentication. It never uses the Control UI preview credential or a human's publication-only My GitHub connection. A configured but unavailable identity fails closed with reconnect guidance. It does not retry anonymously or fall through to another account.

Before saving an HTML or registered widget declaring this binding, the Gateway verifies the selected identity. An unavailable identity or retired caller fails without replacing an existing widget, creating a pending grant, or broadcasting a change. Reconnect the agent's GitHub identity in Settings and retry. Pinning does not query Actions or verify repository permission. Those checks happen on read. Ordinary widgets and MCP App tool names do not trigger this identity check.

Other fields, including identity overrides, URLs, headers, and methods, are rejected. The result keeps GitHub's { total_count, workflow_runs } shape. Each run contains only id, name, display_title, head_branch, status, conclusion, html_url, run_started_at, created_at, updated_at, event, workflow_id, and run_attempt. No credentials or raw repository objects are returned. The upstream response is capped at 1 MiB and the projected run list at 30 entries. Successful reads are cached for about 30 seconds within the current Gateway, board identity, credential, repository, and filter scope.

Rate limits, access denial, unavailable identity, and upstream failure return sanitized guidance. Redirects are refused. For a renamed repository, verify its new name and update both the read and grant. Each caller revalidates its widget, Gateway, and identity before receiving data, including shared reads and cache hits. Removing one widget does not fail another authorized widget's shared read.

Security and storage

Widget documents use restrictive Content Security Policies. Inline style and script are allowed, while arbitrary external resource loads remain blocked. Registered content kinds can load their explicitly public static renderer assets from the isolated sandbox origin. Inline transcript widgets cannot fetch the network. A pinned dashboard widget can fetch only exact HTTPS origins that the agent declared and the session policy granted.

The Control UI's widget content iframe always omits allow-same-origin, even when the global embed mode is trusted, so widget scripts cannot read the parent application origin. With scripts enabled, the outer proxy runs on a dedicated origin and relays messages across the frame boundary. In strict mode, the Control UI still reads the document through its authenticated Gateway connection, but renders it without scripts or scripted interactions. Native clients use isolated, nonpersistent web views and block navigation away from the hosted widget. The core document host also serves widgets with a Content-Security-Policy: sandbox allow-scripts response header, so direct rendering still runs the widget in an opaque origin instead of an application origin. Only render widget code you are willing to execute in that isolated frame.

The iframe also follows gateway.controlUi.embedSandbox. The default scripts tier supports interactive widgets while preserving origin isolation.

The accepted WebRTC data-channel egress residual is documented in Dashboard Architecture.

Canvas retains at most 32 widgets per session (or per agent when no session is available). Creating another widget removes the oldest document in that scope.

Related

• Control UI hosted embeds
• Discord Activities
• macOS widget panel
• Gateway protocol client capabilities