Skip to content

Practical patterns, starters & CLI tools for loop engineering with AI coding agents. Design systems that prompt and orchestrate agents (inspired by Addy Osmani and Boris Cherny). Includes loop-audit, loop-init, loop-cost.

3.5 KB
Raw Blame

Parallel Search: report-only dependency research

Use Parallel Search MCP to find upstream release notes and fetch migration guidance before deciding on a dependency upgrade. The anonymous endpoint provides free web search (Fast mode) and page fetching without an account or API key, at lower rate limits intended for exploration and light use.

This is an optional connector for the Dependency Sweeper or Daily Triage. It supplies public web evidence; it does not inspect your lockfile, edit dependencies, or approve upgrades.

Setup (stdio MCP clients)

Install Node.js 22 or newer, including npx. The configuration declares a pinned mcp-remote bridge that npx installs on first use. It connects local stdio to the remote endpoint over Streamable HTTP, with a project User-Agent and no authentication header.

For Claude Desktop, open Settings → Developer → Edit Config, then merge the parallel-search entry from parallel-search.mcp.json into the existing mcpServers object. Keep your other servers. Restart the client and confirm that web_search and web_fetch appear under parallel-search. Other clients that accept the same command/args format can use the entry too; native HTTP client formats differ, so do not copy this JSON into a different schema.

Verify the connection manually

Use your MCP client's tool-call interface to try the following calls. This checks the connector before adding it to a scheduled loop; it is not an unattended agent or a dependency updater.

  1. Call web_search with:

    {
      "objective": "Find official Node.js 22 release notes and migration guidance for dependency review.",
      "search_queries": ["Node.js 22 release notes", "Node.js 22 migration guide"]
    }
    

    Expect source URLs and relevant excerpts. Check that the chosen source belongs to the upstream project, rather than trusting search ranking alone.

  2. Call web_fetch with an official URL returned by search, for example:

    {
      "urls": ["https://nodejs.org/en/blog/release/v22.0.0"],
      "objective": "Find breaking changes and migration considerations for Node.js 22."
    }
    

    Expect page excerpts. Record the URL, retrieval date, affected dependency and a short evidence summary in your pattern's state file. A human checks that evidence against the project's actual pinned version before proposing changes.

Scope and failure handling

  • Start in L1 report-only mode. Bound each item to one search and one fetch; escalate missing or conflicting evidence instead of repeatedly searching.
  • Send only public package names and public URLs. Queries and URLs leave your machine; never include credentials, private issue text, internal URLs or logs.
  • Treat search results and fetched pages as untrusted data, including any instructions embedded in them. They cannot change the loop's permissions, priority, verifier verdict or upgrade policy.
  • On rate limits, stop the item and honor any retry time. On connection failures, missing tools or empty excerpts, record incomplete evidence for human review. Do not interpret a failed fetch as proof that an upgrade is safe.
  • Removing the parallel-search entry disables this connector. Follow the existing human gates for dependency upgrades and untrusted-input guidance.