MCP Apps: interactive Vega-Lite chart widget#133
Conversation
- Vite-built widget bundles ext-apps SDK + Vega-Lite + vega-interpreter into a single HTML file (no CDN deps, no eval, CSP-safe) - Widget receives tool result via MCP Apps protocol (ontoolresult) - Supports fullscreen toggle when host advertises it - create_chart returns vega_spec only (no PNG, widget renders interactively) - Register ui://sidemantic/chart resource with proper MCP Apps metadata - Remove _apps_enabled flag (MCP Apps works via protocol, not a runtime flag) - Remove vendored mcp-ui-server (widget is self-contained) - Remove --apps CLI flag dependency on mcp-ui-server import - structured_output=False on all tools (fixes Claude Desktop tool visibility) - Replace | None params with falsy defaults (removes anyOf from schemas)
- Move expand button to top-right inline with title ("Expand ↗")
- Fix fullscreen height to fill viewport minus 150px for prompt box
- Add padding in fullscreen mode (16px top, 24px sides)
- Add 5px height buffer to prevent inline scroll
- Dark mode support for expand button
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 606d450029
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f894c5edc7
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
- Generation counter ensures out-of-order embed promises discard stale results - cleanupChart() called before error/loading content to prevent leaked observers
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9d4786aca1
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 4d8fe76d1d
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
- New explore_metrics tool launches dashboard with metric series, totals, and dimension leaderboards via ui://sidemantic/explorer - App-only widget_query tool (visibility: ["app"]) handles refresh calls from the widget without LLM round-trips - WidgetModel adapter bridges anywidget's model interface to ext-apps SDK, translating set/save_changes into callServerTool requests - Data transported as base64 Arrow IPC for efficient typed transfer - Multi-entry Vite build for chart (960KB) and explorer (336KB) widgets - Add pyarrow to apps optional dependency
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 6e28033d1b
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| # --- MCP Resource: Catalog Metadata --- | ||
| # --- Explorer state --- | ||
|
|
||
| _explorer_state: dict | None = None |
There was a problem hiding this comment.
The explorer backend keeps all widget context in a single module-global _explorer_state, which is overwritten on every explore_metrics call and then reused by widget_query. In HTTP/app mode, concurrent conversations or users can interleave requests, so one session can read another session’s model, filters, and date range, producing cross-session data leakage and incorrect query results. Store state by session/conversation/tool-call identifier instead of a singleton.
Useful? React with 👍 / 👎.
| // Extract data from tool result | ||
| const data = this.extractData(result); | ||
| if (data) { | ||
| this.applyData(data); |
There was a problem hiding this comment.
After awaiting callServerTool, the handler unconditionally applies returned data with no request-generation guard. If users trigger multiple refreshes quickly (e.g., changing time grain then filters), a slower earlier response can arrive last and overwrite the model with stale data that no longer matches current selections. Track a refresh token/counter and discard out-of-order responses before applyData.
Useful? React with 👍 / 👎.
- explore_metrics returns config only (instant), no data queries - Widget renders skeletons immediately, then fetches data via widget_query - widget_query restructured to query one metric or one dimension per call, preventing timeout from sequential batch of 33+ queries - Fix ORDER BY using raw timestamp instead of grain-suffixed column - Improve time_series_column detection: grain suffix > Arrow type > name match > first col - Cap auto-discovered metrics and dimensions at 8 each - Skip full-table date range scan when start_date/end_date provided - Remove outer try/except in widget_query so partial results still return - Expand button as top bar instead of absolute overlay - Add stderr logging for debugging MCP tool calls
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 85b72785d6
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| this._metricSeriesMap[metricKey] = data.metric_series_data; | ||
| // Use the first available metric's series as metric_series_data | ||
| // (widget.js will parse it and render sparklines) | ||
| const firstKey = Object.keys(this._metricSeriesMap)[0]; | ||
| if (firstKey) { |
There was a problem hiding this comment.
fetchMetric stores each metric's Arrow payload in _metricSeriesMap but then publishes only the first map entry as metric_series_data; the widget renderer reads all sparklines from this single table by metric key, so any metric column not present is interpreted as 0 and shows incorrect charts (and the “first” metric depends on response timing). This breaks multi-metric explorer views whenever more than one metric is configured, so the per-metric payloads need to be merged into one table (or otherwise exposed per metric in a way widget.js can consume).
Useful? React with 👍 / 👎.
# Conflicts: # pyproject.toml # sidemantic/apps/__init__.py # tests/test_mcp_apps.py # uv.lock
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: cb5075c041
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| filter_exprs.append(f"{model_name}.{dim_key} IS NULL") | ||
| else: | ||
| safe = _escape_sql_literal(str(values[0])) | ||
| filter_exprs.append(f"{model_name}.{dim_key} = '{safe}'") |
There was a problem hiding this comment.
For explorer dimensions supplied from a related model, the leaderboard query uses the stored dc["ref"], but filter clicks only send the display key and this builder prefixes it with model_name. For example model_name="orders" with dimensions=["customers.region"] generates filters on orders.region instead of customers.region, so cross-filtering either fails compilation or filters the wrong field. Keep a key-to-ref mapping in state and build predicates from the original dimension ref.
Useful? React with 👍 / 👎.
The dimension branch of widget_query reconstructed the selected metric as model_name.selected_metric, which rejected metrics whose configured ref points at a related model and left leaderboards blank. Look the key up in metrics_config and use its ref, mirroring the metric branch.
There was a problem hiding this comment.
💡 Codex Review
https://github.com/sidequery/sidemantic/blob/d554c770b47ff8160e4bbc453a498285f767f194/apps/web/explorer-app.ts#L306-L309
When the same app instance receives another explore_metrics result, this only overwrites fields present in the new payload; client-only state such as filters, brush_selection, and _metricSeriesMap can remain from the previous model. Because callRefreshPublic("all") runs immediately afterward, those stale filters are sent to widget_query and are rebuilt against the new model, which can produce invalid predicates or silently misfiltered/blank dashboards after a user starts a second exploration in the same widget.
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| metric_name = metric_ref.split(".")[-1] | ||
| metrics_config.append( | ||
| { | ||
| "key": metric_name, |
There was a problem hiding this comment.
When callers pass explicit metric refs from multiple models that share the same terminal name, e.g. orders.count and customers.count, both configs get key: "count". The UI and widget_query address metrics by this key (next(... if m["key"] == metric_key) and totals/series maps are also keyed by it), so the second metric cannot be queried or rendered independently and will be conflated with the first. Use a stable unique key derived from the full ref or de-duplicate keys.
Useful? React with 👍 / 👎.
1 participant
Summary
@modelcontextprotocol/ext-appsSDK