MCP server
The Vindex MCP (Model Context Protocol) server lets an MCP client call Vindex as a tool. It scores one user story per call and returns the INVEST result with reasoning; it also reports your company’s remaining allowance and explains the dimensions. For the journey content (why, and a walkthrough per client) see Vindex for AI assistants.
Connection details
| Endpoint | https://mcp.vindexgateway.com/mcp |
|---|---|
| Transport | MCP over Streamable HTTP |
| Server name | vindex |
| Sign-in | OAuth 2.1 with PKCE through WorkOS AuthKit, the same sign-in as the Vindex Portal. The client opens your browser once and keeps the session. |
| Who can connect | Any registered Vindex user. New users create their account and company on first sign-in (free plan). No admin action, no key to issue. |
| Company and allowance | Your organization is your company. Scores count against the company's monthly allowance, once per story; re-scores are free. |
The endpoint is the only value you paste into a client. Everything else, including where to sign in, the client discovers from the server. See Authentication for the discovery documents and scopes.
Quick connect
One snippet per client. Each client also has a full setup page with the sign-in step and troubleshooting in its own vocabulary.
Claude Code
Setup pageclaude mcp add --transport http --scope user vindex https://mcp.vindexgateway.com/mcpCodex
Setup pagecodex mcp add vindex --url https://mcp.vindexgateway.com/mcpCursor
Setup page{
"mcpServers": {
"vindex": {
"url": "https://mcp.vindexgateway.com/mcp"
}
}
}GitHub Copilot
Setup page{
"servers": {
"vindex": {
"type": "http",
"url": "https://mcp.vindexgateway.com/mcp"
}
}
}Claude Desktop and claude.ai: Settings, then Connectors, then Add custom connector with the endpoint URL. Setup page.
ChatGPT: Settings, then Connectors, then Create (Developer mode) with the endpoint URL; paid plans. Setup page.
Any other MCP client: add a remote server with the endpoint URL; the client discovers the sign-in.
Every path ends the same way: ask the assistant to score a story with Vindex, or run the score-story prompt.
First call
After sign-in, type a sentence like this into the assistant:
Score this story with Vindex, story id NW-105:
As a customer I want an ETA email when my delivery is delayed.
Acceptance criteria: the customer receives an email with the new ETA; the email
includes the order number and the new window; the customer can opt out in settings.The assistant calls score_story and presents the finding, then the overall percentage and band, then the dimensions on request. Scored through the live server on 2026-09-17, this story returned 62%, Needs review, with Small at 40 and Estimable at 50. To confirm the connection is scoring against the right company, ask the assistant to check Vindex usage; get_usage reports the company name, plan, and allowance.
What the server exposes
Limits and behaviour
| Stories per call | One. There is no batch tool; the assistant scores several stories in sequence and presents each as it arrives. |
|---|---|
| Typical latency | 2 to 10 seconds. The server streams a progress notification every 5 seconds while the engine works. |
| Ceiling | 65 seconds, after which the call returns a tool error: “Vindex did not respond in time. Try again in a moment.” |
| Concurrency | Four in-flight scores per company per server replica. Above that the tool answers “Vindex is busy for your company, try again shortly.” |
| Allowance | Per company, per calendar month (UTC). A story counts once; re-scores are free. Free plan: up to 10 stories a month. When exhausted, new stories are paused until the period ends; re-scores still work. |
| Errors | Returned as tool results with isError true and one plain sentence the assistant can relay. No stack traces, no codes. |
| Health bands | Healthy 80 to 100, Needs review 60 to 79, At risk 0 to 59. Suggestions, not a gate. |
Data handling
Story text is sent to Vindex to be scored and for nothing else. The MCP server is a stateless adapter: it stores no scores, no users, and no keys, and holds only short-lived caches. Scores are stored to your company like every other Vindex source. Vindex does not train models on customer data. See the security page.
Not supported
- Batch scoring. Backlog views belong in the tracker plugins and the Vindex Portal.
- Writing to trackers. The server never writes to Jira, Azure DevOps, or Asana. The plugins own write-back.
- Rewriting stories. Vindex returns reasoning and suggestions. There is no rewrite tool or prompt.
- History and dashboards. Use the Vindex Portal or the Stats API.
Frequently asked
Do I need a separate account for the Vindex MCP server?
No. Sign in with the account you use for the Vindex Portal. If you are new, the sign-in creates your account and your company on the free plan. Every registered Vindex user can connect.
Does connecting cost anything?
No. Connecting is free. Scoring a story counts against your company's monthly allowance once per story; re-scores are free. The free plan covers up to 10 stories a month.
Does a low score block anything?
No. The health band is a suggestion. Vindex returns the band and the reasons; it does not stop, change, or write anything. Any rule that stops a coding agent on a low band is one the team writes into the agent's instructions.
Can my whole team use it?
Yes. Everyone in your organization signs in with their own Vindex account and scores against the company's shared allowance. Users deactivated in Vindex lose access to the tools.
Which clients work?
Any client that supports remote MCP servers over Streamable HTTP with OAuth. Verified: Claude Desktop and claude.ai, Claude Code, ChatGPT, Codex, Cursor, and GitHub Copilot.
Can my coding agent call it?
Yes. Claude Code, Codex, Cursor's Agent, and GitHub Copilot's agent mode and coding agent can all call score_story before they build, and read the structured result.