Skip to Content
The Makinari API repo now includes an MCP Server — connect your AI models directly. View on GitHub →
MCPMedia & research

Media & research tools

MCP names: generateImage, generateVideo, generateAudio, webSearch, urlToMarkdown, urlToSitemap.

generateImage

POST /api/agents/tools/generateImage. Images are stored and returned as URLs.

ParameterTypeNotes
promptstringRequired. Style, color, composition
sizestring256x256, 512x512, 1024x1024 (default)
qualitystringstandard, hd
ratiostring1:1, 4:3, 3:4, 16:9, 9:16, 3:2, 2:3
nnumberCount (default 1)
reference_imagesarrayImage URLs used as context
providerstringGemini only today
POST https://backend.makinari.com/api/agents/tools/generateImage Authorization: Bearer YOUR_API_KEY Content-Type: application/json { "prompt": "A minimalist logo for a tech startup, flat design, blue and white", "size": "1024x1024", "ratio": "1:1" }
{ "success": true, "images": [{ "url": "https://...", "width": 1024, "height": 1024 }] }

generateVideo and generateAudio follow the same auth; audio is TTS from a text prompt.

webSearch

Azure OpenAI Responses with Bing web search. Returns search results and, when requested, an AI answer. Results reflect indexed/cached web data, not live page scraping; source dates are not guaranteed and no relevance score is invented.

ParameterTypeNotes
querystringRequired
search_depthstringbasic (default): up to 1 tool call, low reasoning; advanced: up to 2, medium; comprehensive: up to 3, medium
max_resultsnumberLocal cap 1–20, default 5; does not guarantee this many results from Bing
limitnumberLegacy alias for max_results; if both are supplied they must agree
include_answerbooleanDefault true
include_imagesbooleanDefault false; true is unsupported and rejected
include_raw_contentbooleanRaw content is unsupported; requests to enable it are rejected
date_from, date_to, time_rangestringHard date filters are unsupported and rejected; query wording is not an enforced date range
include_domains, exclude_domainsarrayHost filters
POST /api/agents/tools/webSearch

The depth names are compatibility controls for tool-call budgets, not a guarantee of equivalent Tavily search depth. Commercial credits remain 0.1 per query; Azure provider cost is model input/output tokens plus Bing num_requests.

The API server calls HTTP POST /openai/v1/responses with store: false and must use an existing reasoning-enabled Azure deployment supporting web_search_call.results (gpt-5.2 tested). Credentials use the optional dedicated AZURE_OPENAI_SEARCH_ENDPOINT + AZURE_OPENAI_SEARCH_API_KEY pair first; otherwise MICROSOFT_AZURE_OPENAI_ENDPOINT + MICROSOFT_AZURE_OPENAI_API_KEY if either is set; otherwise AZURE_OPENAI_ENDPOINT + AZURE_OPENAI_API_KEY. Both values of the selected pair are required; incomplete pairs fail rather than mixing or falling back. Deployment precedence is AZURE_OPENAI_SEARCH_DEPLOYMENT, then MICROSOFT_AZURE_OPENAI_DEPLOYMENT, then AZURE_OPENAI_DEPLOYMENT.

Configure Azure secrets in the Vercel API server’s environment, never in client-side NEXT_PUBLIC_* variables. This migration does not edit private .env files or deploy. Bing grounding is a separate compliance boundary: review Bing terms and data-handling requirements; store: false is not a guarantee of Bing retention, residency or Azure model-processing compliance parity.

urlToMarkdown / urlToSitemap

POST /api/agents/tools/urlToMarkdown POST /api/agents/tools/urlToSitemap

urlToMarkdown fetches a URL and returns clean markdown for the model. urlToSitemap walks the site and returns discovered paths.

Last updated on