Dashboards

A dashboard is a durable board of named widgets built by an agent. Use it when people should return to the current output rather than read through a conversation to find the latest result.

Creating a dashboard starts its first build immediately. You provide a title, optional description, instructions, model, reasoning level, and optional project. If no project is chosen, d5s creates an unlisted one for the dashboard.

Refreshable material

Dashboard instructions describe the audience, data to use, questions to answer, and presentation the agent should produce. The agent uses native metric, chart, table, and text widgets for structured results, and custom HTML widgets for interactive views. Each widget has a stable name and can be updated without replacing the rest of the board.

Use Chat, Split, or Dashboard to switch between the conversation and board. Ask the agent to update a named widget in the existing conversation. Use Sources to refresh saved connector datasets independently of the design. Completed agent responses use the same personal feedback controls as chats, including an optional details dialog. The chat panel shows only your own selection, not teammate ratings or aggregate counts. Each published board update records a new board revision.

A manual dashboard refresh starts from the current published board and its existing source bindings. The current widgets, values, notes, and layout are authoritative; the original creation request is not replayed because its literal seed values may be stale. The agent refreshes existing source-backed datasets and changes the board only when fresh data or a new explicit instruction requires it.

The dashboard chat panel also supports /compact for an existing idle build session. Choose it from the compact / menu, then submit again to summarize eligible older conversation context without requesting a dashboard update or creating a chat bubble.

Registered API data sources

Dashboard data sources use supported read operations from a connection available to the dashboard's project. Current connection permissions apply to every refresh. Widgets cannot request arbitrary external destinations or provide credentials.

GitHub pull-request lists are supported for direct API data sources. Connecting an account does not make every API operation available as a dashboard source. MCP sources also require a supported read operation and structured data.

GitHub data sources

A GitHub data source can list open pull requests for a repository using a GitHub connection available to the dashboard's project. Choose the summary fields to publish, such as pull-request number, title, author, and update time.

Refreshing a data source updates its saved data without starting another agent conversation. If access is revoked, the provider fails, or a complete result cannot be collected, the previous complete result is retained.

Publishing an HTML widget

Ask the dashboard agent to publish a saved HTML report as a named widget. The widget appears once its HTML and supporting files have been captured successfully. An existing widget stays visible while its replacement is being prepared.

If the board changes during publication, or the files cannot be captured, the previous widget stays in place. The agent can inspect the current board and submit a new publication when you are ready to try again.

Custom HTML views

A custom HTML widget can use the same saved datasets as native charts and tables. Refreshing a source updates those views without asking the agent to redraw the dashboard. Each source keeps its own refresh state and last successful timestamp.

Navigating an embedded report away from its saved document closes its live data connection.

HTML widgets can follow the app theme and automatically fit their content height, within a bounded preview. Longer reports scroll inside the preview. A frameless presentation lets a report use its own heading and layout and starts at the full board width. Updating a widget preserves your saved arrangement. Narrow boards, including Split view, stack widgets so each remains readable. Source links appear as an Open source link outside the embedded report, where you can open the provider page.

Rename a dashboard

Choose Rename from a dashboard's actions menu, on the Dashboards list, a project's Dashboards strip, or the dashboard page itself. The title is configuration rather than build output, so it stays editable while a build is running and after one fails.

Renaming also retitles the dashboard's build sessions, so the same name appears wherever those sessions are listed. If the dashboard is backed by an unlisted project, that project is renamed too — so promoting it later surfaces the current name rather than the one you first typed. A dashboard inside a project you named yourself never renames that project.

Versions keep history

The History menu contains published board snapshots, including widget content and layout. Preview a snapshot to compare it, or restore it without asking the agent to rebuild the board. Historical HTML views use captured files and do not connect to live datasets.

Reports, CSV datasets, and source code presented in the conversation remain attachments. Opening an attachment shows a separate, closable preview; it never replaces the board. Only an explicit named-widget publication changes dashboard content. A completed agent turn without published widgets is not a live dashboard.

The project supplies reusable instructions, memory, files, skills, connectors, and access. The dashboard owns its instructions, conversation, widgets, saved datasets, layout, and board history. Deleting a dashboard preserves its past build sessions.

Other workspace runs can discover and inspect a dashboard, then contribute by updating its title, description, instructions, or enabled state. Publishing widget content remains an explicit operation in the dashboard conversation.

Native reports

Ask the agent to group an explanation, headline metrics, charts, and supporting tables into one report widget. Reports use the app's own components and move together on the board. Their values are saved snapshots: ask the agent to update the report when its analysis changes. For independently refreshed data, use a dataset-connected metric, chart, or table widget.

Interactive read panels

Custom panels can offer named reads from a saved dataset or a live connector. For example, select an issue status and choose whether to read the saved snapshot or fetch current results. These reads run without an agent rebuilding the panel and do not replace saved dataset snapshots.

The panel can use only its configured sources and permitted filter choices. If the source definition or access changes, the read stops until the panel is updated. Live read permissions must be granted by the source owner, including when restoring an older panel from history.

Work with a workspace agent

Use Work with in dashboard chat to select an accessible workspace agent. Sending a request starts a fresh contribution with that agent's persona and standing instructions and the selected dashboard widget. It does not copy the agent's private conversation. The agent can inspect accessible widgets and datasets as needed.

The selector shows the selected agent’s avatar and persona.

Open Contributions to read a contribution and see its responding agent. When a named contribution ends, the agent’s own conversation receives a compact dashboard card. For people who can still open the dashboard, the card shows the dashboard name, agent, time, the verified outcome, and any widgets known to have updated or failed to publish. A pending HTML publication stays marked as publishing until its result is known. Without dashboard access, the conversation keeps only a generic activity entry and link. The request and contribution transcript are never copied into the agent’s conversation.

A follow-up starts fresh with current dashboard context. Files in a contribution open from that contribution's coding sandbox.

You can also ask an agent from its own chat to inspect and update a dashboard you identify. The original chat stays in its project, and dashboard access is checked separately. If someone changes the same widget while an edit is underway, review its current version before retrying.

Workspace membership does not grant another agent's credentials. Current dashboard, conversation, and agent permissions still apply, as do the selected agent's budget and model policies.

MCP data shapes and limits

MCP data sources can use supported row lists, arrays, or single records. A result may cover only part of the remote account; confirm the source's scope before treating a dashboard as complete.

Refresh requires an allowed, read-only operation. Prose, images, files, and mixed content can inform an agent's analysis but are not automatically refreshable datasets. Verify the first refresh and the displayed data before relying on a live widget.