Do agent-native web search with a remote MCP

Who it's for

For people who want web search for an AI agent and accept aggregating model-native search and agent search providers instead of a traditional search engine. Official copy says Agent-native web search for AI agents — aggregating model-native search and agent search providers, not traditional search engines.

The compatibility list names Codex CLI, Claude Code, OpenCode, Hermes, plus ordinary shell, Python apps, and remote Streamable HTTP MCP clients.

This fits if you can do one remote deploy and accept that search requests leave this machine. Official copy says No telemetry, no shared secrets, and Provider keys stay in runtime environment variables. No shared API-key service does not mean no network, and does not mean no privacy boundary.

Repository: https://github.com/JerryLiu369/agent-web-search

First try in 3 minutes

Use only One-click remote MCP. The official deploy entry is https://vercel.com/new/clone?repository-url=https%3A%2F%2Fgithub.com%2FJerryLiu369%2Fagent-web-search&env=AGENT_WEB_SEARCH_AUTH_TOKEN

Prerequisite: prepare a real token for the environment variable named AGENT_WEB_SEARCH_AUTH_TOKEN; do not fill a placeholder. The page does not write a local npx or pip start command, so do not invent an install line.

One low-risk call: run one search through the native MCP tool. Official copy says a native MCP tool, but does not give a function name; use it as the same provider-neutral search core.

Success signal: you get normalized results. Official docs do not list result fields. model-backed providers may also return answer; the first time, only confirm that results appear.

Remote search is not zero risk. The query goes to the service you deploy and to backend providers.

Use cases

Case 1: Search once with the native MCP tool. After deploy and token setup, let the agent search and check that the return is normalized results. Do not treat the interface as a native API of some traditional search engine.

Case 2: When a model backend is on, also look at answer. Official copy says Every backend is available through the same MCP tool, CLI, Python API, and normalized results; model-backed providers may also return an answer. The second time still uses the same contract; only confirm whether answer appears. The page does not name a CLI executable, so do not invent a command line.

Limits

The same core can also be taught to a CLI through a standard Agent Skill, but the skill file path and command do not appear. Provider keys stay only in runtime environment variables.

The repository title mentions free keyless defaults; the body also requires an auth token and provider keys. Do not deploy without a token. Do not invent a second MCP tool name.