Costex
Open demo Your workspace.
Your tools.
Read Costex from your own software, connect an MCP client, or ask questions inside the application. Each path uses the same scoped cost tools.
Try the sample assistant ↗| Connection | What it gives you | Where it runs |
|---|---|---|
| Read API | Versioned endpoints for cost overview, period comparison, charges, recommendations, resource findings and allocations. | Your private deployment, with a separate machine token. |
| MCP | The same named tools for compatible desktop clients and agents. Streamable HTTP and local stdio transports. | Your private gateway or a local process you control. |
| Assistant | Questions, tool evidence and links back into your workspace. Bring an OpenAI or Anthropic model. | Inside Costex. Provider credentials stay on your server. |
Ask a useful question.
- What changed in our selected reporting period?
- Which services account for the most spend?
- Which resource findings need review?
- How is this cloud and tenant allocated for the billing month?
Evidence before action.
The first integration release is read-only. It cannot delete resources, run SQL, execute agents or create merge requests. Scope, source coverage and tool evidence accompany the results. Provider estimates remain estimates, and a cleanup finding is a review candidate.
Connect a model on your server.
An administrator chooses the provider and model and supplies its API key through private deployment configuration. The user sees that provider before sending a question. Selected cost context is then sent to it. Provider charges and data handling depend on that customer's account; a chat subscription does not configure API access.
About this public preview.
The four suggested questions show prepared examples immediately, with a brief loading transition. Typed questions use the demo's real MCP tools and fictional billing data. Neither calls an AI model, accepts credentials or connects a cloud account. Private API and MCP endpoints are not hosted on this public site.
Client compatibility.
Remote clients need support for a configured bearer header; local clients can launch the stdio server. OAuth discovery and app-directory publishing are not part of this release. A machine token covers one trusted customer workspace. Tenant filters select data for analysis; they do not create separate user permissions.
The private project includes an installation guide, the endpoint contract and executable connection examples. See data and security for the wider deployment boundary.