5 Best Tools to Build Dashboards with Claude Code and Codex

The best tool for building a dashboard with Claude Code or Codex depends on what happens after someone reads the chart. Shortlist ToolJet for operational dashboards with visual editing, Cube for governed business metrics, Retool for teams building within Retool, Streamlit for Python data apps, and Plotly Dash for custom analytical interactions.

A sales report, an inventory console and a forecasting application can all look like dashboards. They have different requirements for data access, editing, deployment and maintenance. This guide explains those differences and gives you a small dataset, a reusable build prompt and concrete checks for evaluating your own result.

Scope and methodology: This is a documentation-based selection guide, reviewed on October 24, 2026. ToolJet publishes this article and is one of the options. Recommendations reflect documented capabilities and the use cases described below, not a new timed benchmark across all five tools. Linked ToolJet examples are separate published builds. The sample exercise is an evaluation brief, not a claim that every product passed it.

Tool	Best fit in this shortlist	How the agent helps; main trade-off
ToolJet	Operational dashboards that combine charts, tables and business actions	MCP builds native ToolJet apps; authorized builders can continue visually. SQL, permissions and custom logic still need review.
Cube	Consistent metrics across business reports and analytical dashboards	MCP and agent skills work with its semantic layer and dashboards. Define the business metrics and access policies first.
Retool	Teams that want agent-assisted apps inside their existing Retool environment	Its CLI supports agent-written app code; its MCP has a different building and credit model. Choose the route deliberately.
Streamlit	Python-based exploration and lightweight analytical apps	A coding agent writes and edits Python files. A technical owner maintains the code and deployment.
Plotly Dash	Bespoke analytical applications with detailed chart interactions	A coding agent writes layouts and callbacks. Flexibility comes with code ownership and operational work.

These are different categories, not interchangeable products ranked by a single score. Start with the outcome, then test the most relevant two options. If you want a broader platform comparison without a coding-agent requirement, see our low-code dashboard builders guide.

Claude Code and Codex can connect to external tools through MCP. That connection exposes particular operations; it does not mean every connected product can create a dashboard. A connector that only reads data is different from one that creates pages, edits queries or publishes an application. Check the supported operations before choosing your stack. See the official Claude Code MCP documentation and OpenAI MCP documentation.

For a code-based framework, the agent can instead work on the project files. In that route, the framework does not need a special dashboard-generation connector: someone still needs to run, test and deploy the resulting application. Do not confuse that with a vendor-supported MCP integration.

Claude chat is also a different starting point from Claude Code. Confirm that a tutorial supports the client you actually use, along with its installation method and permissions. This article is about creating business dashboards, not dashboards that monitor token usage or coding-agent activity.

ToolJet is a strong fit when the dashboard is part of daily work: inspect an overdue order, assign an owner, update a record or open a detail page. Its MCP integration lets Claude Code or Codex build native ToolJet apps that remain editable in the visual builder. App builds through this route use the coding agent’s subscription rather than ToolJet AI credits.

That makes the handoff worth testing. Ask the intended maintainer to rename a chart, move a filter and change a table column after generation. Visual changes can be accessible to authorized builders; query logic, SQL and server-side access rules still require someone who understands them.

Use the ToolJet MCP setup guide for the current installation instructions. Our hospital analytics dashboard example shows a KPI row, chart grid and language selector backed by ToolJet DB. It is an example of the output, not a guarantee that your first prompt will produce the same result.

Check before choosing: required connectors, deployment model and plan entitlements. Keep agent usage separate from ToolJet licensing and infrastructure. A self-hosted dashboard does not by itself keep prompts or retrieved data away from the external model provider.

Cube is worth evaluating when the difficult part is agreeing on what “revenue,” “active customer” or “retention” means across reports. Its official agent-skills repository documents Claude Code and Codex support, a hosted MCP connection, semantic-model work and dashboard creation. The connection applies the signed-in user’s access rules. Source: Cube agent skills.

Choose this route when reusable metric definitions matter more than adding operational forms to the same screen. During evaluation, ask two dashboards for the same metric and check that their totals, filters and time zones agree. Do not let the model silently invent a new definition because the prompt was vague.

Check before choosing: who owns the semantic model, which data sources are available, viewer and embedding requirements, and the approval process for model changes. Cube lists its MCP server and dashboard capabilities on its pricing page; compare the plan and compute requirements for your workload.

Retool documents both CLI and MCP routes for use with Claude Code or Codex. Its current guidance recommends the CLI for building and maintaining apps, while MCP is suited to tasks such as exploring resources and managing the instance. In the documented building comparison, CLI work uses the local agent’s credits; the MCP building route can also consume Retool credits. Source: Retool’s building-surface comparison.

This is a practical candidate if your organization already maintains Retool resources and wants to build within that environment. Ask which app-building experience the generated output uses and how another team member will edit it. Do not assume every agent-generated app behaves like a classic drag-and-drop app.

Check before choosing: the chosen route’s availability, resource access, review workflow and total licensing. Our ToolJet and Retool hands-on comparison provides separate historical test evidence; its individual runs are not performance estimates for this dashboard exercise.

Streamlit is a sensible candidate when the person maintaining the dashboard already works in Python. Its official tutorial demonstrates loading and caching data, drawing charts and maps, and using widgets to filter results. A coding agent can help write that application, but the deliverable remains a Python project. Source: Streamlit app tutorial.

Use it for an analyst’s exploratory dashboard or an interface around an existing Python calculation. Evaluate the maintenance handoff honestly: if the next person cannot run the project or review a changed calculation, generating the first version has not solved ownership.

Check before choosing: hosting, package dependencies and data authorization. Streamlit supports OIDC authentication, but authentication alone does not decide which records a person should see. Implement and test those restrictions separately. Source: Streamlit authentication documentation.

Plotly Dash is worth considering when a dashboard needs deliberately designed interactions: selecting a point updates another chart, a control changes a calculation, or users compare scenarios. Dash describes the UI through its layout and connects inputs to outputs through callbacks. A coding agent can help implement those files. Source: Dash callbacks documentation.

The attraction is control over the analytical experience. The corresponding responsibility is maintaining the code, dependencies and deployment. Choose it when a developer or data scientist will own the app, rather than assuming that a generated codebase becomes a no-code product.

Check before choosing: callback behavior, query cost, responsiveness and the deployment’s access controls. Test your expected data volume and concurrent use. A smooth demonstration with six rows says little about a large production dataset.

Use fictional data before connecting real business systems. The following six-row order dataset makes mistakes easy to spot. All amounts are USD and represent order value, not accounting revenue. Cancelled orders are excluded from the included order-value metric.

order_id,order_date,region,status,amount_usd
O101,2026-10-01,West,Completed,100
O102,2026-10-02,East,Completed,200
O103,2026-10-03,West,Pending,150
O104,2026-10-04,East,Cancelled,80
O105,2026-10-05,West,Completed,250
O106,2026-10-06,East,Pending,120

Paste this brief into the agent after setting up the chosen platform. Replace the bracketed platform name. Treat any request to deploy, expose data or change permissions as a separate decision.

Build a sales operations dashboard in [platform] using only the fictional CSV supplied.

Show included order value, included order count, pending order count, a daily order-value chart and an order table.
Exclude Cancelled orders from included value and count.
Add region and date filters. Filters must apply consistently to every metric, chart and table.
Use USD labels and date-only values without shifting dates between time zones.
Provide loading, empty and error states. Show when the data was last refreshed.
Keep this first version read-only. Do not connect production systems or publish it.

Before building, explain the data model and metric definitions.
After building, show the queries or calculations and verify the expected totals below.
Identify anything that requires manual setup or could not be verified.
Do not describe the app as production-ready.

Check	Expected result
All regions, October 1–6	Included value $820; 5 included orders; 2 pending orders
West only, October 1–6	Included value $500; 3 included orders; 1 pending order
East only, October 1–6	Included value $320; 2 included orders; 1 pending order
All regions, October 1–3 inclusive	Included value $450; 3 included orders; 1 pending order
Date range outside the dataset	Zero totals and a clear empty state, not stale figures

These are arithmetic expectations from the sample rows, not observed product results. Independently check the totals before accepting the agent’s summary. Then change one row and refresh the dashboard: the display should reflect the new value rather than a hard-coded answer.

The most useful follow-up prompt is a small change with a measurable result: “Add a status filter without changing the existing metric definitions, then repeat the all-regions and West-only checks.” Watch whether the agent preserves the working dashboard and explains what changed.

Next, ask a colleague to make a visual adjustment without the original conversation. Can they find the right chart, inspect its data source and revert the edit? For a code project, can they run the app from the repository instructions? For a visual platform, can they distinguish layout changes from query and permission changes?

Record the time spent correcting errors and handing over the result, not only generation time. A ten-minute first draft that nobody can safely maintain may cost more than a slower build with clear ownership.

Area	Evidence to require
Metric accuracy	A named owner for each definition; totals reconciled with a trusted query or report.
Data access	An allowed-user and denied-user test against server-side permissions, including exports and direct requests.
Agent access	A scoped connection for the builder; no assumption that its permissions equal the dashboard viewer’s.
Model approval	The exact approved provider, model and endpoint, plus a review of what prompts and tool results leave your environment.
Operations	Refresh schedule, failure state, logs, rollback procedure and an owner for incidents.
Usability	Keyboard navigation, readable chart labels, mobile layout and an understandable empty state.
Writes and actions	Separate validation and authorization tests before enabling record edits, approvals or workflow triggers.

Start with read-only access to a small approved dataset. Add writes only when the read path and permission checks are correct. Hiding a button or applying a browser filter is not a substitute for enforcing access on the server. Our secure internal dashboards guide covers this evaluation in more detail.

Use four separate lines in your estimate: agent usage, platform licensing, runtime infrastructure and maintenance. Include the initial build, corrections and a representative monthly change. For hosted tools, check viewer, builder, embedding and environment charges where applicable. For code projects, include hosting, authentication, monitoring and developer ownership.

If the dashboard contains an AI assistant, its production inference cost is another line item. The model used to build the dashboard does not automatically pay for the AI features used by every viewer. A cheap generation run is therefore useful evidence about that run, not the dashboard’s annual cost.

For an operations team that needs charts alongside business actions and wants a visual maintenance path, start with ToolJet. For a data team standardizing metrics, evaluate Cube. If Retool is already your application environment, test its CLI route. If Python ownership is established, compare Streamlit with Dash based on the interaction complexity you actually need.

Use one dataset and the same acceptance checks for your shortlist. The right result is a dashboard your team can trust, operate and change after the agent conversation ends. Start with the ToolJet MCP overview or the setup guide if that is the route you want to evaluate.