search-boost-mcp

search-boost-mcp

A multi-engine web search MCP server that integrates with coding agents, offering parallel search, page fetching, X/Twitter search, deep research, and configurable free/API layers.

Category
访问服务器

README

search-boost-mcp

Multi-engine web search MCP server for coding agents. One CLI install wires it into Cursor, Cursor CLI, Codex, Claude Code, Grok Build, and Antigravity.

search-boost family

Project For Link
search-boost (this repo) Cursor · Codex · Claude · Grok · Antigravity via MCP you are here
dsh-search-boost DeepSeek Harness bundle plugin GitHub · npm
pi-search-boost pi extension GitHub · npm

Search engines are vendored in lib/search/ (originally from dsh-search-boost): on the free layer, Bing, DuckDuckGo, Yahoo, and Exa-free run in parallel; the api layer adds Antigravity CLI (when available) and whichever keyed Tavily / Brave / Exa engines you configure (one key is enough; all three recommended for best fusion). Also included: X/Twitter fallback, Jina page fetch, and deep-research rounds.

中文文档 → README_zh.md


Install (recommended)

Requires Node ≥ 22.13.

npm install -g search-boost-mcp
search-boost setup          # interactive: keys → layer → agents
# or non-interactive:
search-boost install -y     # all detected agents

Restart each agent after install so MCP reloads.

Upgrade

Already using search-boost-mcp? Update the global CLI, then refresh agent configs (keys/layer under ~/.search-boost/ are kept):

npm install -g search-boost-mcp@latest
search-boost install -y                 # all detected agents
# or pick targets, e.g. Grok one-step (plugin + config when grok on PATH):
search-boost install -t grok -y --auto-allow
search-boost doctor

Restart each agent after reinstall. Config lazy-migrates from older flat ~/.search-boost-*.json paths on first write.

One-liners by agent

search-boost install -t cursor -y
search-boost install -t codex,claude -y --auto-allow
search-boost install -t grok -y --auto-allow   # plugin + config when grok CLI on PATH
search-boost install -t antigravity --workspace --auto-allow -y

Preview without writing: search-boost install --dry-run -y

Verify install

search-boost doctor          # health checks (offline, pass/warn/fail)
search-boost doctor --json   # machine-readable report for CI/scripts
search-boost status          # install dashboard (keys, layer, agents)

Exit codes: 0 healthy · 1 failure (or warnings with --strict) · 2 warnings only.

Optional: search-boost doctor --probe adds live search smoke (needs network; Phase 2).

Then confirm in your agent:

Agent Quick check
Cursor Settings → MCP → search-boost connected; tools listed
Cursor CLI Same MCP entry in ~/.cursor/mcp.json
Codex codex session lists mcp__search-boost__* tools
Claude Code MCP panel shows search-boost; tools callable without deny prompt (if --auto-allow)
Grok Build grok mcp doctor search-boost · grok inspect
Antigravity MCP config includes search-boost; restart IDE after install

If tools appear but calls fail, run search-boost serve in a terminal to see startup errors.

Install flags (quick reference)

Flag Effect
-t, --target Which agent(s) to wire (cursor, codex, claude, grok, antigravity, cursor-cli, auto, all)
-y, --yes Non-interactive: skips keys/layer wizard, uses --target=auto, implies --auto-allow and --replace-native
-t without -y Still non-interactive for that target and still replaces native search by default — but does not imply --auto-allow; add it explicitly if you want no permission prompts
--auto-allow Pre-approve search-boost MCP tools in agent config (Cursor CLI allowlist, Codex default_tools_approval_mode, Claude/Grok/Antigravity permission rules) so the agent does not prompt every session
--replace-native / --keep-native Disable or keep built-in web search where the agent supports a switch (Codex web_search, Claude WebSearch). Default is replace when non-interactive
--scope user|project|all Grok only: user (~/.grok), project (.grok/ in cwd), or both on uninstall
--skip-grok-plugin Grok only: skip bundled grok plugin install; still writes config.toml, rule, and skill
--dry-run Print actions without writing files

For full onboarding (API keys + layer choice), run search-boost setup or search-boost install without -y.

Uninstall

search-boost uninstall -t codex -y
search-boost uninstall -t cursor,codex,claude -y

Uninstall removes only search-boost-owned blocks (marked MCP entries, skills, hooks, permission rules). Where the agent supports it, native web search is restored (Codex top-level web_search, Claude WebSearch deny) unless you used --keep-native at install time or had pre-existing unmarked settings. Config files created solely for search-boost are unlinked when empty after cleanup. For Grok, plugin removal via grok plugin uninstall is best-effort (warns and continues if the CLI is missing or uninstall fails). Preview: --dry-run.


What you get

MCP tool Purpose
fused_search Multi-engine parallel search, dedupe, cross-ranking
fetch_page Full page text (Jina + HTML fallback, optional focus)
x_search X/Twitter keyword / user / thread
deep_research One round per call — repeat with suggested_queries until gaps empty, then synthesize (~3 rounds max)
search_layer Show or set free (keyless) vs api (keyed engines)
search_stats Cache hits, engine availability, diagnostics

Also: resource search-boost://policy · prompt search_routing

Layers

  • free — Bing + DuckDuckGo + Yahoo + Exa-free; no API keys.
  • api — free-layer engines plus Antigravity CLI (when available) and any of Tavily / Brave / Exa that you configure (one key works; all three recommended for best cross-engine fusion)

Keys: search-boost config keys~/.search-boost/config/keys.json (flat ~/.search-boost-keys.json and legacy ~/.dsh-search-boost-keys.json still read; or env TAVILY_API_KEY, BRAVE_API_KEY, EXA_API_KEY). Optional routing: enabledEngines: ["exa"] or "engines": { "brave": { "enabled": false } } in the keys file.

Config layout: runtime data lives under ~/.search-boost/config/ (keys, layer, xauth), cache/ (xguest token), state/ (Antigravity workspace registry). First write lazy-migrates from flat ~/.search-boost-*.json and legacy ~/.dsh-* files (old copies kept). Override base: SEARCH_BOOST_HOME; per-file: SEARCH_BOOST_*_FILE.

Obtain keys: Tavily · Brave Search API · Exa

X/Twitter auth (optional): improves official x_search when credentials are present. Stored at ~/.search-boost/config/xauth.json (flat/legacy paths still read) or via XAI_API_KEY. Configure with search-boost config x (see CLI cheat sheet). MCP /x-login and search-boost config x write the same local copy. Override file path: SEARCH_BOOST_XAUTH_FILE.

search-boost config x --show              # xauth status
search-boost config x --import-grok       # import grok CLI login
search-boost config x --set-xai-key KEY   # store XAI API key
search-boost config x --logout            # remove local copy

Config file overrides: SEARCH_BOOST_KEYS_FILE, SEARCH_BOOST_LAYER_FILE, SEARCH_BOOST_XAUTH_FILE (optional env vars pointing at custom paths).


CLI cheat sheet

Command What it does
search-boost Interactive TUI
search-boost setup Onboarding (keys + layer + install)
search-boost install / uninstall Wire MCP + prompts into agents
search-boost serve Run MCP stdio server (used by agents)
search-boost status Keys, layer, X credentials, per-agent configured state
search-boost doctor [--quick|--probe] [--json] [--strict] Config/agents/engine health checks with pass/warn/fail
search-boost config keys|layer|x|search Keys, default layer, X auth, native-search replace
search-boost print <agent> Print MCP snippet without writing
search-boost agents Machine-readable agent list

Install flags: -t cursor,codex,…|auto|all · -y (non-interactive; implies --auto-allow + --replace-native) · --dry-run · --auto-allow (pre-approve MCP tools — see table above) · --replace-native / --keep-native · --scope user|project|all (Grok) · --skip-grok-plugin (Grok) · --workspace (Antigravity .agents/)


Supported agents

Agent MCP config Also installs
Cursor IDE ~/.cursor/mcp.json hook, skill
Cursor CLI ~/.cursor/mcp.json (same surface as IDE) hook, skill (CLI variant), optional CLI auto-allow
Codex CLI ~/.codex/config.toml AGENTS.md, skill
Claude Code ~/.claude.json CLAUDE.md, skill, permissions
Grok Build ~/.grok/config.toml rule, skill, bundled grok-plugin (when grok on PATH)
Antigravity ~/.gemini/config/mcp_config.json AGENTS.md, GEMINI.md, skill, optional workspace

Prompts use model-discretion wording (search when you choose — not forced every turn). See agents/ for per-agent templates.

Native web search: With --replace-native (default when non-interactive), Codex gets a marked top-level web_search = "disabled" in config.toml (never inside [mcp_servers.*]); Claude gets an ownership-marked WebSearch deny in settings.json. Uninstall removes only search-boost-owned entries and restores native search when safe. Cursor / Antigravity rely on skill + hook preference only. Grok native browse is left on.

Cursor + Cursor CLI: Both targets share one ~/.cursor/ surface. Installing -t cursor,cursor-cli merges IDE + CLI prompts into a single write; uninstall clears the shared surface.

Antigravity + agy CLI: On the api layer, the optional Antigravity CLI engine (agy on PATH) joins medium and complex fused_search tiers only — not simple lookups. It depends on local sign-in and platform quota; timeouts are ~45s.

Grok Build: search-boost install -t grok -y --auto-allow runs grok plugin install <bundled grok-plugin> --trust when the Grok CLI is on PATH, then writes config.toml, rule, and skill. If grok is not on PATH, the plugin step is skipped with a warning and the config install still proceeds. Use --skip-grok-plugin for config/rule/skill only. Re-install is idempotent for [permission] blocks (marked or legacy); uninstall strips search-boost-owned permission lines. If [ui] permission_mode = "always-approve", --auto-allow skips injecting [permission]. The plugin's .mcp.json uses portable npx; config.toml uses resolveMcpLaunch() (local node when developing from a clone) — both can coexist. Manual plugin install: grok plugin install ./grok-plugin --trust (advanced) → grok-plugin/README.md.


Troubleshooting

Symptom What to try
Is search-boost healthy? search-boost doctor — pass/warn/fail verdict; --json for scripts
Install fails immediately Node ≥ 22.13 (node -v); upgrade if older
MCP server missing in agent Re-run install, restart the agent, check search-boost status
Tool calls blocked / approval every turn Re-install with --auto-allow, or approve once in the agent UI
No results / empty engines search-boost doctor — check layer/keys/engine checks; free needs no keys; api needs at least one keyed engine via search-boost config keys or env vars (all three recommended)
Network/proxy issues Phase 2: search-boost doctor --probe (not yet implemented)
MCP won't start search-boost doctormcp_launch_command, node_version; then search-boost serve
Grok plugin MCP won't start grok mcp doctor search-boost; ensure npx and network access work
grok fails on config.toml parse Duplicate [permission] — run search-boost install -t grok -y --auto-allow
Antigravity agy never runs Requires api layer, agy on PATH, and complexity medium/complex — not simple
Timeouts / fetch errors Corporate proxy or firewall may block Bing/DDG/Jina; try search-boost serve locally to read stderr

Development

git clone https://github.com/Mr-remon219/search-boost.git
cd search-boost && npm install
npm run check && npm run test:install && npm run smoke
node cli.mjs install --dry-run -y

Local clone installs write node /path/to/cli.mjs serve (not npx). No sibling checkout or SEARCH_BOOST_DSH_ROOT is required.


License

MIT


Links: Issues · dsh-search-boost · pi-search-boost

Friendly link: LINUX DO 社区

推荐服务器

Baidu Map

Baidu Map

百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。

官方
精选
JavaScript
Playwright MCP Server

Playwright MCP Server

一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。

官方
精选
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。

官方
精选
本地
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。

官方
精选
本地
TypeScript
VeyraX

VeyraX

一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。

官方
精选
本地
graphlit-mcp-server

graphlit-mcp-server

模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。

官方
精选
TypeScript
Kagi MCP Server

Kagi MCP Server

一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。

官方
精选
Python
e2b-mcp-server

e2b-mcp-server

使用 MCP 通过 e2b 运行代码。

官方
精选
Neon MCP Server

Neon MCP Server

用于与 Neon 管理 API 和数据库交互的 MCP 服务器

官方
精选
Exa MCP Server

Exa MCP Server

模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。

官方
精选