본문으로 건너뛰기
C

modsearch-plus

ciskonc/modsearch-plus

Fork of liustack/modsearch: a web-search bridge for text-only agents that adds MiniMax engines and a Zhipu engine (coding-plan MCP or balance REST, one engine with a channel setting) to a parallel multi-engine chain with per-engine failover, health lights and a Plugins-page settings card. Its HTTP routes have no DSH access check, so a web page open in your browser could call them.

설치

dsh plugin --profile web add github:ciskonc/modsearch-plus

README

modsearch-plus

modsearch-plus

Web search, X search, and page fetch for chat apps whose models cannot search the web — free, no signup, no API key to start.

Engines run as a chain with automatic failover: keyless engines are in place by default, keyed engines take over as quota and budget allow.

GitHub license GitHub last commit GitHub issues GitHub contributors npm version

English | 中文


Introduction

modsearch-plus is a web search plugin for chat applications. It gives the model web search, X search, and page fetch behind a single engine chain that fails over automatically, and it is free to start: no signup, no API key. The repository is a fork of liustack/modsearch, based on upstream v5.10.10 with the 5.10.11 and 5.10.12 upstream changes merged in, and extends the chain with native MiniMax and Zhipu search engines.

A chat model with no search tool can only answer from what its training data holds. News published after its cutoff and pages that are still changing are out of its reach, and topics it was never trained on are more hopeless still. Faced with those, the model guesses or admits it does not know.

Once the engine chain is wired in, search sits behind the chat itself: a question that needs the web triggers a search mid-conversation.

Screenshot

The settings card in the DeepSeek Harness web app: pick a preferred engine, enter API keys, set a self-hosted endpoint, and tick the engines that join the failover chain.

Settings card

Search and fetch as the model sees them:

Web search in DeepSeek Harness Search in Codex

Engines

Engines run as a chain and fail over automatically: when one is rate-limited, out of quota, or fails authentication, the next engine takes the request. Firecrawl is the keyless default, so search works before any configuration.

EngineCapabilityFree quotaActivation
FirecrawlWeb search + page fetch1,000 free credits/monthDefault, keyless
KeenableWeb search1,000 requests/hour/IPKeyless
You.comWeb search100 requests/day/IPKeyless
Antigravity CLIWeb search + page fetchFreeInstall the CLI and sign in
TavilyWeb search1,000 credits/monthAPI key
ExaWeb search$10/month of credits (about 1,400 searches, no card needed)API key
MiniMaxWeb searchM Plan subscription keyAPI key
ZhipuWeb search (MCP or REST)Coding-plan credits, or pay-per-use 0.01 CNY/request on the balanceAPI key; the channel setting picks the transport
MiniMax websearchWeb searchPay-per-use, 0.03 CNY/requestAPI key
Grok BuildX (Twitter) searchRides a SuperGrok or X Premium subscriptionInstall Grok Build and sign in
localPage fetchFreeNo setup

Keyed engines accept comma-separated multiple keys — Tavily, Exa, Firecrawl, Keenable, MiniMax, MiniMax websearch, and Zhipu — and rotate them automatically on authentication failures, rate limits, or exhausted quota. Keys live in ~/.modsearch/config.json (permissions 0600) or in environment variables such as TAVILY_API_KEY. Self-hosted endpoints are set per engine, for example modsearch config set tavily.baseURL <url>.

Installation

DeepSeek Harness (recommended)
npx -y @deepseek-ai/dsh plugin --profile web add modsearch-plus@5.11.0

In the desktop app, enter modsearch-plus in the Add plugin dialog instead. On dsh releases before 0.1.6-alpha.2, the settings card lives under Settings → Plugins → Plugin config.

Other harnesses

Claude Code, Codex, Pi, and OpenCode install modsearch as a skill. The skill itself comes from the upstream project:

npx skills add liustack/modsearch
Build from source
git clone https://github.com/ciskonc/modsearch-plus.git

Usage

There are no commands to memorize. After install, the search ability is simply there: when a question needs the web, the model searches mid-conversation, and the answer carries its sources. Aggregated content ships with an uncertainty field marking how reliable it is.

Settings card

In the DeepSeek Harness settings card you pick a preferred engine, enter API keys, set a self-hosted endpoint, and tick the engines that join the failover chain. Saving takes effect immediately, and keys already stored are never sent to the browser.

How it plugs into DSH

The engine chain becomes the search provider of DSH's native web seam — the built-in web_search tool runs on modsearch — while x_search and read_page are registered as raw tools the model can call directly.

Compatibility

Plugin versionDSH version testedNotes
5.11.0 (2026-11-21)0.2.2-rc.5The two Zhipu engines merge into one zhipu engine with a channel setting (auto / coding / rest); the settings card lays the chain out as one draggable column. Breaking: glm-coding.* and glm.* settings are not migrated, set the key as zhipu.apiKey
5.10.15 (2026-11-20)0.2.2-rc.5Repository metadata now points at this fork; screenshots.json added
5.10.140.2.2-rc.5Settings card fix for DSH rc.5
5.10.xDSH ≥ 0.2.xThe 5.10.x line is tested on DSH 0.2.x and later

Non-DSH hosts (Claude Code, Codex, Pi, OpenCode) run modsearch as a skill, so the DSH column does not apply there.

Credits

modsearch-plus is a fork of liustack/modsearch: the engine chain and the plugin itself come from upstream, and this fork adds the native MiniMax and Zhipu engines on top. Thanks to the upstream project.

Thanks to the engine providers behind the chain: Firecrawl, Keenable, You.com, Antigravity CLI, Tavily, Exa, MiniMax, Zhipu, and Grok.

License

MIT. Licensing information is maintained in the repository.

관련 플러그인