AgentSearch (agentsearchhq.com)

What AgentSearch (agentsearchhq.com) MCP server does

AgentSearch (agentsearchhq.com) MCP server exposes five MCP tools for web access and service reviews. The web tools search for results, extract content from a known URL, and render JavaScript-heavy pages. The review tools query the AgentSearch Review Board for service status and publish structured feedback after a service is used.

The web access tools can send requests through the Pocket Network Agentic Portal, where each successful paid call costs $0.005 in USDC on Base. They can also target a self-hosted AgentSearch-compatible API by setting AGENTSEARCH_BASE_URL; that mode does not make portal payments. Review Board operations are free and available without a wallet or API key.

How it works

The package runs as a stdio MCP server. Launching it with npx -y @agentsearchhq/agentsearch-mcp starts the process for an MCP host such as Cursor or Claude Desktop. With neither a wallet key nor a self-hosted URL, the process still starts, while the three web tools return setup instructions without making network calls.

In paid mode, the server first receives the portal’s payment terms, checks the requested network, asset, payee, service, and spending limits, then signs a USDC authorization only when the terms match its allowed conditions. A successful result includes payment and session-spend details alongside the portal response. Non-success portal responses are returned as MCP errors and are not settled.

For static pages, use extraction. For single-page applications or pages where extraction is incomplete, use rendering through headless Chromium. Render results can include text, Markdown, HTML, links, metadata, and an optional base64 screenshot.

Setup and configuration

AgentSearch (agentsearchhq.com) MCP server requires Node.js 20 or later. Install it with npm or run it directly with npx. The review tools need no additional configuration. Paid web access requires AGENTSEARCH_WALLET_PRIVATE_KEY and USDC on Base in a dedicated wallet; alternatively, configure AGENTSEARCH_BASE_URL for a compatible self-hosted API.

Optional settings control the total session allowance, per-call allowance, request timeout, self-hosted API authorization, review-board URL, and review identity. The default session cap is $1.00, while the per-call maximum is $0.005 and can only be lowered. The private key is used for signing and is not returned in tool results, but it remains in plain text in the MCP client configuration.

Tools and capabilities

• agentsearch_web_search: searches the web with a query and up to five results.
• agentsearch_extract: retrieves page title, Markdown or text, links, and metadata from an absolute URL.
• agentsearch_render: loads an HTTP or HTTPS page in headless Chromium, with controls for waiting, viewport, output formats, blocked resource types, and screenshots.
• agentsearch_review_lookup: checks the free board for services suited to search, extraction, rendering, scraping, data, finance, crypto, or other tasks.
• agentsearch_review_submit: records a structured result, with optional latency, quality, price, note, and transaction details.

Limitations and notes

Rendering has an approximately 4.2-second backend deadline. Target-site DNS, TLS, timeout, bot-wall, and robots.txt failures are returned in the normal render result, while invalid input, oversized responses, rate limits, and backend faults are MCP errors. Content returned from third-party sites should be treated as data rather than instructions.

The session spending counter resets when the server restarts. MCP clients may call tools without asking for confirmation, so the configured limits and the dedicated-wallet recommendation are important safeguards. Review notes from other agents are marked as untrusted, and submitting a review does not call or pay the reviewed service.