Connect Ristria to Claude Code, Cursor & OpenCode
The Ristria MCP connector runs locally and forwards each tool call to the managed Ristria cloud. Provide your RISTRIA_API_KEY; nothing else to install or host on your machine.
// ~/.config/opencode/opencode.jsonc or claude_desktop_config.json
{
"mcpServers": {
"ristria": {
"command": "npx",
"args": ["-y", "@ristria/mcp"],
"env": {
"RISTRIA_API_KEY": "rs_free_YOUR_API_KEY",
"RISTRIA_API_URL": "https://api.ristria.com"
}
}
}
}Save 3,000 System-Prompt Tokens with the Ristria CLI
Coding agents like Claude Code, OpenCode, and Aider can invoke the ristria binary directly from Bash. This eliminates MCP tool-schema overhead at rest while preserving full JSON output and --max-tokens context packing.
export RISTRIA_API_KEY="rs_free_YOUR_API_KEY"
export RISTRIA_API_URL="https://api.ristria.com"
# 1. Search with a token budget cap
npx -y -p @ristria/mcp ristria search "Python asyncio TaskGroup exception handling" --limit 5 --max-tokens 600
# 2. Extract HTML, /llms.txt, or PDF documents
npx -y -p @ristria/mcp ristria extract "https://docs.python.org/3/library/asyncio-task.html" --max-chars 8000
# 3. Map a documentation site (sitemaps + link discovery)
npx -y -p @ristria/mcp ristria map "https://modelcontextprotocol.io" --limit 20
# 4. Find externally similar technical pages
npx -y -p @ristria/mcp ristria similar "https://docs.python.org/3/library/asyncio-task.html" --limit 5
# 5. Estimate monthly workload cost across providers
npx -y -p @ristria/mcp ristria cost --searches 25000 --extracts 5000 --research 500Call Ristria from Any Language: POST /v1/search
Building with LangChain, LlamaIndex, or custom Python/TypeScript code? Send a Bearer request to /v1/search with your RISTRIA_API_KEY. Highlights and max_tokens packing are included for 1 credit.
curl -X POST https://api.ristria.com/v1/search \
-H "Authorization: Bearer rs_free_YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"query": "Google Search Central robots.txt documentazione italiano",
"category": "general",
"include_highlights": true,
"max_tokens": 600
}'All 12 Tools & REST Endpoints
| MCP TOOL | REST ROUTE | CREDITS | CAPABILITY |
|---|---|---|---|
| growth_search_cascade | POST /v1/search | 1 cr | Web search with rank fusion, Okapi BM25 + embeddings reranking, recency and entity-overlap boosts, fast-path highlights, and max_tokens packing. |
| growth_search_social | POST /v1/search (site:) | 1 cr | Targeted vertical search across X (x.com), Reddit, GitHub, and Hacker News. |
| growth_find_similar | POST /v1/similar | 1 cr | Profiles a seed URL and ranks externally related pages using lexical, entity-overlap and dense-embedding similarity. |
| growth_map | POST /v1/map | 1 cr | Discovers site URLs via /robots.txt, /sitemap.xml (<sitemapindex>), concurrent link discovery, and a headless fallback for JavaScript SPAs. |
| growth_crawl | MCP stdio / Engine | 5 cr | Combines growth_map discovery with 3-worker parallel extraction and query-driven BM25 passage selection. |
| growth_research | MCP stdio / Engine | 3 cr (mini) / 5 cr (pro) | 2-hop agentic research: runs sub-queries in parallel, derives emergent technical terms for Hop 2, fuses via RRF, and generates cited reports. |
| growth_extract_content | POST /v1/extract | 1 cr | Extracts clean Markdown/text from HTML, /llms.txt, .md, and PDF documents, with a 7-day cache. |
| growth_extract_batch | MCP stdio / Engine | 2–3 cr | Parallel extraction for up to 20 URLs with optional query-driven BM25 highlight selection. |
| growth_extract_structured | MCP stdio / Engine | 1 cr | DOM table and row extraction with deterministic where filtering (e.g. price > 100 && name ~ /^pro/i). |
| growth_analyze_content_gap | MCP stdio / Engine | 1 cr | Calculates a deterministic domain-diversity score (0–100) across top SERP results for SEO/AEO topic analysis. |
| growth_run_discovery | MCP stdio / Engine | 3 cr | Monitors 10 default AI/engineering sources and persists latest-report.json. |
| growth_get_latest | MCP stdio / Engine | 1 cr | Reads the cached discovery report from disk with zero network calls. |
Anti-SSRF Protection on Every Fetch
Every URL fetched by extract, map, crawl, or similar is resolved and validated before connection and on every redirect hop (up to 5 hops), blocking loopback, link-local, and RFC1918 private IPv4/IPv6 ranges.