> ## Documentation Index
> Fetch the complete documentation index at: https://docs.praxa.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Connect Praxa MCP to Claude, Cursor, VS Code, and OpenAI

> Configure a deployment-specific Praxa remote MCP server in common AI hosts, preserve OAuth and approval boundaries, and verify tool discovery safely.

Connect hosts to the canonical Streamable HTTP `/mcp` URL supplied by your
Praxa Integration Gateway deployment. The examples use
`https://gateway.example.com/mcp` as a placeholder; replace it only with an
approved deployment URL. A successful configuration entry does not prove OAuth,
scope, tenant policy, or tool execution.

## Before you configure a host

1. Obtain the exact HTTPS endpoint and supported OAuth flow from the deployment owner.
2. Inventory the `aura_*` tools the workflow actually needs.
3. Separate safe reads from mutations and destructive cancellation.
4. Decide which calls require human approval in the host.
5. Use a disposable user and tenant for initial testing.

## Claude Code

```bash theme={null}
claude mcp add --transport http praxa https://gateway.example.com/mcp
claude mcp list
claude mcp get praxa
```

Open Claude Code and run `/mcp` to complete OAuth when the server requests it.
Use project scope only when the shared configuration contains no credential.
Review the server before accepting project trust.

## Cursor

Create `.cursor/mcp.json` for a project-scoped remote server:

```json theme={null}
{
  "mcpServers": {
    "praxa": {
      "url": "https://gateway.example.com/mcp"
    }
  }
}
```

Use Cursor's MCP settings or `cursor-agent mcp login praxa` for OAuth, then run
`cursor-agent mcp list-tools praxa`. Keep auto-run disabled for mutations until
approval, replay, and negative authorization tests pass.

## Visual Studio Code

Create `.vscode/mcp.json`:

```json theme={null}
{
  "servers": {
    "praxa": {
      "type": "http",
      "url": "https://gateway.example.com/mcp"
    }
  }
}
```

Run **MCP: List Servers**, start `praxa`, review the trust prompt, and inspect
the MCP output log. Do not hardcode OAuth tokens in a committed workspace file.
Enterprise administrators should allowlist the exact server and preserve tool
approval policy.

## OpenAI Responses API

Your trusted backend can provide a short-lived delegated token to the remote
MCP tool. Keep approval required and allowlist only the tools needed for the
request.

```typescript theme={null}
import OpenAI from "openai";

const openai = new OpenAI();
const response = await openai.responses.create({
  model: "gpt-5.6",
  input: "Read the governed reference coverage and summarize only the evidence.",
  tools: [{
    type: "mcp",
    server_label: "praxa",
    server_description: "Governed Praxa Integration Gateway tools",
    server_url: process.env.PRAXA_MCP_URL!,
    authorization: process.env.PRAXA_DELEGATED_TOKEN!,
    allowed_tools: ["aura_get_coverage"],
    require_approval: "always",
  }],
});

console.log(response.output_text);
```

The application owns the OAuth flow and token lifecycle. Data sent through a
remote MCP tool is also subject to that server deployment's retention and
residency policy.

## Any other MCP 2025-11-25 host

Configure one Streamable HTTP endpoint and negotiate the lifecycle before
listing or calling tools. On JSON-RPC `POST` requests, advertise both
`application/json` and `text/event-stream`; on a listening `GET`, advertise
`text/event-stream`. Preserve the negotiated `MCP-Protocol-Version` on later
requests. If initialization returns `MCP-Session-Id`, send it on every later
request, start a new initialization after a session-scoped `404`, and terminate
the session with `DELETE` when the server supports it. Use OAuth Protected
Resource Metadata discovery instead of a static bearer header unless the
deployment explicitly documents another mode.

## Verify every host

| Check            | Expected result                                                                                |
| ---------------- | ---------------------------------------------------------------------------------------------- |
| Server identity  | `aura-agent-os`, software version `0.3.0`                                                      |
| Protocol         | `2025-11-25` or negotiated `2025-03-26`                                                        |
| Session          | Returned `MCP-Session-Id` is preserved, rejected across clients, and reinitialized after `404` |
| Tool list        | 12 exact `aura_*` definitions                                                                  |
| Schema           | Draft 2020-12 object schemas and original bounds                                               |
| Read-only call   | Correct result with matching scope                                                             |
| Wrong-scope call | Fail closed without data                                                                       |
| Mutation         | Host approval plus stable idempotency key                                                      |
| Destructive call | Explicit approval and post-call readback                                                       |
| Revocation       | Later call fails after token revocation                                                        |
| Logging          | No OAuth token or sensitive payload                                                            |

For exact commands and UI paths, follow the host's current official MCP guide;
host configuration formats can evolve independently of Praxa contracts.

## Official host references

* [Claude Code MCP guide](https://docs.anthropic.com/en/docs/claude-code/mcp)
* [Cursor MCP guide](https://docs.cursor.com/context/model-context-protocol)
* [Visual Studio Code MCP servers](https://code.visualstudio.com/docs/copilot/chat/mcp-servers)
* [OpenAI connectors and remote MCP tools](https://developers.openai.com/api/docs/guides/tools-connectors-mcp)