MCP setup

Two transports, three steps, one minute. Start with Streamable HTTP for production and stdio for desktop dev.

Use a sk_sandbox_* or sk_test_* key. The server validates your key on startup and exits if it’s invalid or revoked.

Streamable HTTP (recommended)

No local install. Add to your client config:

{
  "mcpServers": {
    "ovra": {
      "type": "streamable-http",
      "url": "https://api.getovra.com/api/mcp",
      "headers": {
        "Authorization": "Bearer sk_sandbox_..."
      }
    }
  }
}

Stdio (local subprocess)

Runs the server via npx — no global install required.

{
  "mcpServers": {
    "ovra": {
      "command": "npx",
      "args": ["@ovra/mcp@latest"],
      "env": {
        "OVRA_API_KEY": "sk_sandbox_..."
      }
    }
  }
}

Where the config lives

Client	Path
Claude Desktop (macOS)	`~/Library/Application Support/Claude/claude_desktop_config.json`
Claude Desktop (Windows)	`%APPDATA%\Claude\claude_desktop_config.json`
Claude Desktop (Linux)	`~/.config/Claude/claude_desktop_config.json`
Cursor	`~/.cursor/mcp.json`
Other	See your client docs

Environment variables

Variable	Required	Default	Description
`OVRA_API_KEY`	Yes (stdio)	–	`sk_sandbox_` / `sk_test_` / `sk_live_*` (live in v1.3+)
`OVRA_AGENT_TOKEN`	Optional	–	Use an `at_*` agent-scoped token instead of an API key
`OVRA_API_URL`	No	`https://api.getovra.com`	Override base URL

You can use either OVRA_API_KEY (org-wide) or OVRA_AGENT_TOKEN (single-agent). Agent tokens give the LLM strict isolation — it sees only that one agent’s cards, intents, and transactions.

Verify

After restart your client should expose 19 ovra_* tools. Smoke-test with:

List my agents using ovra_agent action "get".

If the tool list is empty:

Check the Authorization header / OVRA_API_KEY env value is set correctly.
Confirm the key isn’t revoked (GET /keys returns the active list).
Tail your client’s MCP logs — auth failures surface as 401 from the bootstrap call.

Startup sequence

Read credentials

Server reads OVRA_API_KEY or OVRA_AGENT_TOKEN from env (stdio) or Authorization header (HTTP).

Health check

Boot calls GET /health with the credential to validate.

Valid credential → register all 19 tools with their schemas. Invalid → exit with a clear error.

Ready

Client begins discovering tools.

Security in MCP

Every tool response goes through redact() before reaching the model.
Destructive actions require confirm: true.
IDs are sanitized against path traversal.
Webhook and execute URLs must be HTTPS and SSRF-safe.

Tool reference

Every action and parameter.

MCP overview

Architecture, security, supported clients.

Claude Desktop

Step-by-step desktop setup.

MCP Server

Streamable HTTP (recommended)

Stdio (local subprocess)

Where the config lives

Environment variables

Verify

Startup sequence

Security in MCP

Next

Tool reference

MCP overview

Claude Desktop

​Streamable HTTP (recommended)

​Stdio (local subprocess)

​Where the config lives

​Environment variables

​Verify

​Startup sequence

​Security in MCP

​Next

Tool reference

MCP overview

Claude Desktop

Streamable HTTP (recommended)

Stdio (local subprocess)

Where the config lives

Environment variables

Verify

Startup sequence

Security in MCP

Next