MCP server

Nexevo exposes a Model Context Protocol (MCP) server, so an MCP client — Claude Desktop, Cursor, or any MCP-capable host — can call a safe subset of your workspace's tools, including searching your own knowledge.

Endpoint

A single JSON-RPC 2.0 endpoint (MCP revision 2025-11-25, streamable HTTP):

POSThttps://nexevo.ai/v1/mcp

Authentication

Use the same workspace API key as the REST API, in the Authorization header. The key binds the connection to one workspace — a client only ever sees that workspace's tools and data.

http

Authorization: Bearer sk-ws-...

What's exposed

Only a safe, read / compute-only subset of the agent toolset is available over MCP (call tools/list to see the exact list for your workspace):

Web research — web_search, url_reader, advanced_scrape, document_read, ocr_extract.
Open data — arxiv_search, sec_edgar, world_bank, oecd, financial_data, currency_convert, weather, rss_read.
Your workspace knowledge — rag_search over its long-term memory and uploaded knowledge, plus spreadsheet_analyze.
Compute — calculator, convert, current_time, base64, random_choice.

Every tool carries MCP annotations — all safe-subset tools are marked readOnlyHint: true, destructiveHint: false, idempotentHint: true, so clients (Claude Desktop, Cursor) can label them correctly. Tools that return structured data (e.g. rag_search) also declare an outputSchema, and tools/call includes a structuredContent field alongside the text — clients that prefer machine-readable output can consume it directly.

Generation (images / video / speech), customer-service actions, email, the browser / desktop, connectors, and SQL are not exposed over MCP — so a connected client can read and reason, but cannot spend heavily or change anything.

Methods

Standard MCP: initialize (with version negotiation — send your desired protocolVersion in params; we support 2025-11-25, 2025-06-18, and 2025-03-26, echoing back yours or falling back to our latest), the notifications/initialized notification, tools/list, tools/call, and ping.

Try it with curl

List the available tools:

bash

curl https://nexevo.ai/v1/mcp \
  -H "Authorization: Bearer $NEXEVO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Call one — search your workspace knowledge:

bash

curl https://nexevo.ai/v1/mcp \
  -H "Authorization: Bearer $NEXEVO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call",
       "params":{"name":"rag_search","arguments":{"query":"our refund policy"}}}'

Connecting an MCP client

In any MCP host that supports a remote streamable-HTTP server with custom headers, point it at the endpoint and pass your key. The shape varies by client, but the two things every client needs are the URL and the Authorization header:

json

{
  "mcpServers": {
    "nexevo": {
      "url": "https://nexevo.ai/v1/mcp",
      "headers": { "Authorization": "Bearer sk-ws-..." }
    }
  }
}

Tool calls over MCP are billed the same per-call rate as the agent uses (most read-only tools are free). Responses are sanitized — an MCP client never sees raw internal errors.