Give your agent the web.
Eight MCP tools, one API key. Reads are free. Writes charge credits. Errors are typed. Your agent never blocks more than 60 seconds.
Selenium glue, no memory.
LLM agents need the web. The current options are: a brittle Selenium wrapper somebody glued behind an HTTP endpoint, a “search API” with no scrape, or a “scraping API” with no search. None of them speak MCP. None of them remember what the agent did yesterday.
Eight MCP tools. Reads are free.
Stekpad ships an MCP server with eight tools: scrape, crawl, map, extract, search, plus three read tools for the agent’s stored datasets (list_datasets, get_dataset, query_dataset). The reads are free — the agent should never hit the paywall to remember its own work.
Research and remember.
Search for “vector databases with hybrid search 2026”, scrape the top 5 results, save them to a dataset called “Vector DB research”, and summarize each in one paragraph.
Claude calls:
- 1.
searchwithscrape_results: true,num_results: 5, dataset attached — costs 5 + 5 = 10 credits. - 2.
ai_summaryover the 5 new rows — costs 5 credits.
Total: 15 credits (~5 cents). The dataset persists. Tomorrow:
Show me the Vector DB research dataset and re-rank by support for hybrid search.
Claude calls query_dataset (free) and reasons over the rows.
Built for the loop.
Never blocks for 60s
Long crawls return a run_id and the tool yields. The model decides when to check.
Typed errors
target_blocked, insufficient_credits, session_unavailable — each with a message and a retry_after.
Per-call credit reporting
Every response includes credits_charged so the agent can budget itself across a long task.
Free dataset reads
list_datasets, get_dataset, query_dataset are 0 credits. Memory should never have a paywall.
Claude Desktop config.
{ "mcpServers": { "stekpad": { "command": "npx", "args": ["-y", "@stekpad/mcp"], "env": { "STEKPAD_API_KEY": "stkpd_live_..." } } }}Cursor / Claude Code — see /integrations/cursor and /docs/mcp.
What an agent loop costs in credits.
| Operation | MCP tool | Credits |
|---|---|---|
| Search top 5 + scrape | search | 10 |
| Summarize 5 rows | ai_summary | 5 |
| Map a competitor's site | map | 1 per 1k URLs |
| Scrape 50 pages | crawl | 50 |
| Read a dataset | query_dataset | 0 |
Common questions.
Will the agent burn my credits accidentally?
Set a `credits.low` webhook threshold (default 1,000) and Stekpad warns you before the wallet hits 0. Hard-stop at 0, no overage.
Can two agents share a dataset?
Yes — they share a workspace. Use a separate API key per agent if you want isolation in audit logs.
Does the MCP server work with Claude Connect?
Yes. Use the hosted HTTP MCP at `https://mcp.stekpad.com`.
Install MCP for Claude.
300 free credits a month. Reads are always free.