dashies-plugin

<div align="center"> # Dashies **Publish AI-built dashboards to a shareable URL, then let them refresh themselves on a schedule - with no AI in the loop after the first build.** A one-install plugin for [Claude Code](https://claude.com/claude-code), [Codex](https://developers.openai.com/codex), and [Cursor](https://cursor.com): the `dashies` authoring skill plus the Dashies publish MCP server, bundled together. [Website](https://dashies.xyz) · [Marketplace](https://dashies.xyz/marketplace.json) · [MCP server](https://mcp.dashies.xyz/mcp) · [Report an issue](https://github.com/Dashies-ai/dashies-plugin/issues)       </div> --- ## What it is Dashies turns a Claude Code conversation into a published, self-contained BI dashboard at a stable URL: `https://dashies.xyz/<your-handle>/<slug>`. You describe the dashboard. Claude builds it as a single HTML file (inline CSS, inline JS, embedded data) and publishes it with one MCP call. You get back a link you can share. The part that makes Dashies different: **a dashboard can keep itself up to date.** When it is backed by a connected data source, Claude authors the dashboard plus a small JSON manifest (a cube of SQL, dimensions, and additive measures) exactly once. From then on a server-side cron re-runs that SQL on your chosen schedule and rewrites only the numbers in the dashboard's data island. The HTML is never touched again, and no model runs again. > [!NOTE] > Anyone can publish a dashboard. Scheduled auto-refresh is the flagship feature, and it requires the dashboard to be backed by a connected data source. A one-off static dashboard publishes fine - it just will not refresh on its own. ## Install ### One command (recommended) ```sh npx plugins add Dashies-ai/dashies-plugin ``` This is the command on [dashies.xyz](https://dashies.xyz). The vendor-neutral [`plugins`](https://github.com/vercel-labs/open-plugin-spec) CLI auto-detects the AI coding agents you have installed and sets up Dashies in each - the `dashies` authoring skill plus the publish MCP - in one step. It installs into **Claude Code** and **Cursor** out of the box. For **Codex** specifically, or on **Windows**, use the matching per-tool command below - the `plugins` CLI's Codex install and Windows agent-detection are still maturing. ### Claude Code Two commands inside Claude Code: ```text /plugin marketplace add https://dashies.xyz/marketplace.json /plugin install dashies@dashies ``` ### Codex Two commands in your shell: ```sh codex plugin marketplace add Dashies-ai/dashies-plugin codex plugin add dashies@dashies ``` If the marketplace clone fails on SSH host-key verification, use the explicit HTTPS URL: `codex plugin marketplace add https://github.com/Dashies-ai/dashies-plugin.git`. ### Cursor Install Dashies from the [Cursor marketplace](https://cursor.com/marketplace) (Cursor 2.5+) - search **Dashies** under Customize -> Plugins, or run `/add-plugin` in the editor. _(Marketplace listing pending Cursor's review; until it lands, the one-click MCP below works today.)_ In every case, the plugin bundles the `dashies` authoring skill and wires up the Dashies MCP server, so there is nothing to configure by hand and no token to paste. > [!TIP] > This replaces the older manual MCP setup (`claude mcp add dashies ...` or `codex mcp add dashies --url ...`). If you ran that before, the plugin install supersedes it. ### Authentication: no tokens, ever Dashies uses the standard MCP **OAuth 2.1 + PKCE** flow with **Dynamic Client Registration**. There are no API keys or tokens to copy, paste, or rotate. The first time you use a Dashies MCP tool, a browser opens for a one-click Google sign-in. The resulting access token is stored in your OS keychain and refreshes automatically; when it expires, the next tool call re-runs the one-click handshake. ### Using another AI tool? This plugin makes Claude Code, Codex, and Cursor one-install paths, but Dashies is a plain remote MCP server, so any MCP-capable client can publish to it. Point your tool at the same URL - `https://mcp.dashies.xyz/mcp`: - **Cursor (MCP only, no skill)** - prefer the one-install plugin above. For just the MCP, one-click **Add to Cursor** on [dashies.xyz](https://dashies.xyz), or add `{ "dashies": { "url": "https://mcp.dashies.xyz/mcp" } }` under `mcpServers` in `~/.cursor/mcp.json`. - **Codex (MCP only, no skill)** - prefer the one-install plugin above. For just the MCP, run `codex mcp add dashies --url https://mcp.dashies.xyz/mcp`, then `codex mcp login dashies`. - **Claude web, desktop, and Cowork** - add a custom connector for that URL under Settings -> Connectors. The same one-click OAuth sign-in applies everywhere, with no token to paste. [dashies.xyz](https://dashies.xyz) has copy-paste setup for each tool. ## Quick start Once installed, just ask: ```text Publish a dashboard of last month's signups by plan. ``` Claude builds a self-contained HTML dashboard, calls `publish_dashboard`, and hands back a live URL: ```text https://dashies.xyz/<your-handle>/signups-by-plan ``` That is the full loop: one request, one shareable link. ## How the no-AI refresh works The model runs once, at authoring time. After that, refresh is pure server-side cron: it re-executes the saved SQL and writes the new numbers back into the published file's data island. The HTML, the layout, and the bindings stay untouched. ```mermaid flowchart LR subgraph once["Authored once (AI in the loop)"] A["You describe the dashboard"] --> B["Claude builds HTML + manifest:<br/>cube SQL, dimensions,<br/>additive measures"] B --> C["publish_dashboard"] end C --> D[("dashies.xyz/your-handle/slug")] subgraph loop["On schedule (no AI)"] E["Cron fires:<br/>hourly / daily / weekly / monthly"] --> F["Re-run the saved cube SQL<br/>against your connected source"] F --> G["Write fresh numbers into<br/>the data island only"] end G --> D ``` Why this design holds up: - **The HTML never changes after publish.** Refresh swaps data, not markup, so a dashboard cannot drift or break visually between runs. - **No model is in the refresh path.** Lower cost, deterministic output, and no risk of an AI re-interpreting your dashboard differently each cycle. - **The cube is the contract.** A low-cardinality grain plus additive measures is what lets the same SQL re-aggregate forever without re-authoring. ## Features - **Publish in one call.** A self-contained HTML dashboard becomes a stable, shareable URL scoped to your handle. - **Self-refreshing dashboards.** Pick hourly, daily, weekly, or monthly. The server re-runs your cube SQL and updates the numbers, with no AI in the loop. - **Zero-token auth.** OAuth 2.1 + PKCE + Dynamic Client Registration. One browser click, keychain-stored, auto-refreshing. - **Safe renames.** Renaming a dashboard preserves the old URL with a 301 redirect, so links you already shared keep working. - **Versioned bodies.** Republishing snapshots the prior body automatically (up to 20 retained). List and roll back to any snapshot. - **Sandboxed by design.** Published dashboards render under a strict sandbox CSP, isolated from the rest of the origin. - **Guided authoring.** The bundled `dashies` skill walks you through the connected-data gate, cube design, the data-island binding contract, the sandbox constraints, schedule choice, and the manifest. ## What's bundled ### The `dashies` authoring skill Loads automatically after install. It is the playbook for building a *refreshable* dashboard: the connected-database gate, designing a cube with a low-cardinality grain and additive measures, the data-island and data-dash binding contract, the inlined client runtime, the sandbox CSP constraints, choosing a schedule, and writing the `source_config` manifest. ### The MCP tools OAuth triggers on first use of any tool below (one browser click). | Tool | What it does | |---|---| | `publish_dashboard` | Upload a self-contained HTML / JSON / CSV / image file (up to ~5 MB) and get a stable URL back. | | `update_dashboard` | Edit metadata (name, tags, chart, visibility) or rename the slug without re-uploading the body. Renames keep the old URL alive via a 301 redirect. | | `get_dashboard` | Read back a previously published file. | | `delete_dashboard` | Retire a dashboard by slug. The URL stops resolving and the bytes are removed. | | `list_dashboards` | Enumerate your active dashboards, newest first, with cursor pagination. | | `list_dashboard_versions` | List the saved prior body snapshots of one of your personal dashboards, each with a `version_id`. | | `restore_dashboard_version` | Roll a personal dashboard back to a prior snapshot. The current body is snapshotted first, so the restore is normally reversible. | | `get_dashboard_version` | Read back the body of a prior snapshot of a personal dashboard, to inspect or diff it before restoring. | | `update_dashboard_version` | Set or clear a snapshot's label (a short name like "Before redesign"). Metadata only; the body is untouched. | | `introspect_schema` | Inspect a connected data source's schema while authoring a cube - the built-in `self` metrics view, or a warehouse connection you own. | | `validate_cube_sql` | Check a cube's SQL against a connection (`self` or a warehouse you own) before publishing a refreshable dashboard. | | `list_connections` | List the warehouse connections you own - Postgres, BigQuery, Snowflake, or Amazon Redshift (id, label, engine, status). Read-only, never returns secrets; warehouses are connected in the Dashies web app. | | `get_refresh_status` | Check whether a personal dashboard is refreshing on schedule: cadence, next run, last run, and recent run history. Read-only. | | `get_source_config` | Read back the stored refresh manifest (`source_config`) of a personal dashboard, exactly as saved. Read-only. | ## Build your first refreshable dashboard A refreshable dashboard needs a connected data source, so the authoring flow starts there. 1. **Connect a data source.** Auto-refresh re-runs SQL against a live source, so this is the gate. Connect a warehouse - a Postgres database, a BigQuery project, a Snowflake account, or an Amazon Redshift warehouse - in the Dashies web app (the Connections page), credentials go through the app, never the AI - or build against Dashies' own built-in `self` metrics. Without a source you can still publish a static dashboard, but it will not refresh on its own. 2. **Ask Claude to build it.** For example: *"Build a refreshable dashboard of weekly active users by plan, and refresh it daily."* The `dashies` skill takes over from here. 3. **Design the cube.** Claude finds your connection with `list_connections`, uses `introspect_schema` to read its tables, then defines a cube: a low-cardinality grain (the dimensions you slice by) and additive measures (counts, sums) that can be re-aggregated safely every cycle. 4. **Validate the SQL.** `validate_cube_sql` confirms the cube runs against your source before anything is published. 5. **Publish with a schedule.** Claude calls `publish_dashboard` with the HTML plus the `source_config` manifest and your chosen cadence (hourly / daily / weekly / monthly). You get back `https://dashies.xyz/<your-handle>/<slug>`. 6. **Walk away.** The cron re-runs the cube SQL on schedule and writes fresh numbers into the data island. The dashboard stays current with no further AI involvement. > [!TIP] > For a one-off snapshot with no live data behind it, you do not need the cube or the manifest. Just ask Claude to publish a static dashboard - one `publish_dashboard` call and you have a URL. <details> <summary>Common follow-up requests</summary> - **"What dashboards do I have?"** - Claude calls `list_dashboards` and summarizes them, newest first. - **"Rename this dashboard."** - Claude calls `update_dashboard` with a new slug. The old URL 301-redirects to the new one, so shared links keep working. - **"Roll this back."** - Claude calls `list_dashboard_versions` to find a snapshot, then `restore_dashboard_version` to restore it (personal dashboards only). - **"Make it private."** - Claude calls `update_dashboard` with the visibility change; no re-upload needed. </details> ## Links - Home: https://dashies.xyz - Marketplace manifest: https://dashies.xyz/marketplace.json - MCP server: https://mcp.dashies.xyz/mcp - Issues: https://github.com/Dashies-ai/dashies-plugin/issues ## License [MIT](LICENSE). --- <div align="center"> Built by [Dashies](https://dashies.xyz). Dashboards that keep themselves up to date. </div>