Wiro MCP Server

Official MCP server for Wiro AI: search, price and run image, video, audio and LLM models from Cursor, Claude Code, Claude Desktop, Windsurf and any other MCP client.

Website · Models · Docs · Dashboard

@wiro-ai/wiro-mcp is a TypeScript Model Context Protocol server that runs locally over stdio with npx. Its 13 tools let an AI assistant search the Wiro AI model catalog, read a model's parameters and price, run it and wait for the result, upload input files, browse past tasks and search the Wiro AI docs. The hosted server at https://mcp.wiro.ai/v1 serves the same tools with nothing to install.

Quick Start

Local npx and library usage require Node.js 20 or later. The hosted MCP endpoint does not require a local Node.js installation.

1. Get API Keys

Sign up at wiro.ai and create a project at wiro.ai/panel/project/new to get your API key and secret.

2. Add to Your AI Assistant

Cursor: open MCP settings (Cmd+Shift+P → "Open MCP settings") and add:

{
  "mcpServers": {
    "wiro": {
      "command": "npx",
      "args": ["-y", "@wiro-ai/wiro-mcp"],
      "env": {
        "WIRO_API_KEY": "your-api-key",
        "WIRO_API_SECRET": "your-api-secret"
      }
    }
  }
}

Claude Code:

claude mcp add wiro -- npx -y @wiro-ai/wiro-mcp

Then set environment variables WIRO_API_KEY and WIRO_API_SECRET.

Claude Desktop: add to claude_desktop_config.json:

{
  "mcpServers": {
    "wiro": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@wiro-ai/wiro-mcp"],
      "env": {
        "WIRO_API_KEY": "your-api-key",
        "WIRO_API_SECRET": "your-api-secret"
      }
    }
  }
}

Windsurf: add to MCP settings:

{
  "mcpServers": {
    "wiro": {
      "command": "npx",
      "args": ["-y", "@wiro-ai/wiro-mcp"],
      "env": {
        "WIRO_API_KEY": "your-api-key",
        "WIRO_API_SECRET": "your-api-secret"
      }
    }
  }
}

OpenClaw: configure a remote server under mcp.servers:

{
  "mcp": {
    "servers": {
      "wiro": {
        "url": "https://mcp.wiro.ai/v1",
        "transport": "streamable-http",
        "connectionTimeoutMs": 10000,
        "headers": {
          "Authorization": "Bearer ${WIRO_MCP_AUTH}"
        }
      }
    }
  }
}

Set WIRO_MCP_AUTH in OpenClaw's environment to your API key, or to API_KEY:API_SECRET for a project with signature-based auth. Then verify with openclaw mcp doctor wiro --probe. On older OpenClaw releases without that probe command, use openclaw mcp show wiro, restart OpenClaw, and run /tools verbose.

Hermes: add to ~/.hermes/config.yaml:

mcp_servers:
  wiro:
    url: "https://mcp.wiro.ai/v1"
    headers:
      Authorization: "Bearer ${WIRO_MCP_AUTH}"
    timeout: 60
    connect_timeout: 10

Store WIRO_MCP_AUTH (your API key, or API_KEY:API_SECRET) in ~/.hermes/.env, then run /reload-mcp and /tools.

3. Start Using

Ask your AI assistant:

• "Generate a photorealistic image of a mountain lake at sunset"
• "What video generation models are available on Wiro?"
• "Show me the parameters for google/veo3-1"
• "Create a 5-second video with Kling V3: a drone shot over mountains"
• "Check the status of my last task"
• "Show my latest productions and open the newest result"

Authentication

Wiro AI supports two authentication types, chosen when you create a project:

Signature-Based (Recommended)

More secure. Requires both API key and API secret.

WIRO_API_KEY=your-api-key
WIRO_API_SECRET=your-api-secret

API Key Only

Simpler. Only requires the API key. Omit WIRO_API_SECRET from your config.

WIRO_API_KEY=your-api-key

The hosted server takes the same credentials as a bearer token: Authorization: Bearer API_KEY:API_SECRET for a signature-based project, or Authorization: Bearer API_KEY for an API Key Only project.

Available Tools

For long generations, run_model returns the existing task ID and token when its wait budget expires. The assistant must continue with wait_for_task; retrying run_model creates a separate billable task.

LLM-friendly responses

All tools publish typed MCP input and output schemas. Successful calls return structuredContent for reliable chaining, concise text content for clients that only render text, and MCP resource links for generated media. Model parameters remain catalog-driven: json and json-array fields are sent as structured values in the same run_model.params object as all other inputs. Use get_model_schema as the source of truth for continuation fields. previousTaskToken is the single continuation reference. Send it with exactly one new prompt or non-empty messages value for a stateful next turn, or with toolOutputs for a tool-result continuation.

LLM task outputs preserve the public Task Detail segment names and casing: thinking, answer, function_call, and custom_tool_call, including call_id, finishreason, and optional normalized token usage. Only validated fields are projected into structured MCP output. When segments are present, response is derived from ordered answer segments instead of raw debug transport; tasks without a typed projection retain their normal debugoutput response.

Completed tool calls return a continuation template in the existing nextAction. Execute each call, replace the output placeholders, and pass the same ordinary run_model.params object back:

{
  "tool": "run_model",
  "arguments": {
    "model": "owner/model",
    "params": {
      "previousTaskToken": "task-token",
      "toolOutputs": [
        {
          "call_id": "call_01",
          "output": "<tool result>"
        }
      ]
    }
  }
}

Function calls carry JSON-string arguments; custom tool calls carry free-form input. Tool results are request values in toolOutputs, not response segments.

Completed media responses also include an assistant-audience delivery instruction. Some MCP clients expose a resource_link as tool metadata instead of previewing it inside the tool card, so the assistant is explicitly told to render images and preserve clickable media links in its user-facing response. The standard resource link and full-resolution URL remain available to every client; no base64 media payload is required.

Task responses use the stable states submitted, running, completed, failed, and cancelled. When more work is required, the response includes an executable nextAction:

{
  "state": "running",
  "task": {
    "id": "123",
    "token": "task-token",
    "status": "task_start"
  },
  "outputs": [],
  "nextAction": {
    "tool": "wait_for_task",
    "arguments": { "tasktoken": "task-token" },
    "reason": "Continue this exact task. Do not call run_model again."
  }
}

Generation chain: search_models → get_model_schema → run_model → wait_for_task

Cross-conversation history: list_tasks → get_task

Hosted MCP Server

Wiro AI also hosts the same tools at https://mcp.wiro.ai/v1 over streamable HTTP, so nothing has to be installed locally. Send your credentials in the Authorization header as shown under Authentication. The MCP Server documentation has setup instructions for each client.

Install as a plugin

wiroai/Wiro-Plugin packages the hosted server as a plugin for Cursor, Claude Code, Gemini CLI and Devin. Its README has the install command for each of them.

Documentation

• MCP Server (Hosted): setup guides for Cursor, Claude Code, Claude Desktop, Windsurf, Gemini CLI, Devin, OpenClaw, and Hermes
• Self-Hosted MCP: run locally with npx, environment variables, library usage
• Node.js Library: use WiroClient from this package in your own Node.js or TypeScript code
• Authentication: signature-based vs API Key Only
• Run a Model: how the Run endpoint works
• Tasks: task lifecycle, statuses, determining success
• Concurrency Limits: concurrent task limits based on balance
• Error Reference: error codes and handling

Using as a Library

This package exports its core components for use in custom MCP servers:

import { WiroClient } from '@wiro-ai/wiro-mcp/client';
import { createMcpServer } from '@wiro-ai/wiro-mcp/server';

const client = new WiroClient('your-api-key', 'your-api-secret');
const server = createMcpServer(client);

If a project uses signature-based auth and no API secret was given, tool errors say that the secret is missing. When your server takes credentials in its own way, tell users where the secret goes:

const client = new WiroClient(apiKey, apiSecret, undefined, {
  missingSecretHint: 'This project uses signature-based auth. Enter API_KEY:API_SECRET as the API key.',
});

Exports

// Import specific components
import { WiroClient } from '@wiro-ai/wiro-mcp/client';
import { registerTools } from '@wiro-ai/wiro-mcp/tools';

Environment Variables

More from Wiro AI

Website · Models · Docs · X

License

MIT, see LICENSE.