@vcxmug/dsh-native-web

Native (no-MCP, no-cloud) web tools for DeepSeek Harness agents.

The built-in web search depends on a cloud provider (extra token cost), and MCP bridges add protocol hops and session turns. This package instead talks directly over HTTP to your local Firecrawl-compatible web instance — one fetch per call, no bridge, no cloud dependency.

Tool	Purpose
native_search	real-time web search (query, limit, lang/country)
native_scrape	one URL → clean markdown or raw HTML

Requirements

• Node 18+ (global fetch).
• A reachable web instance — typically self-hosted Firecrawl (docker compose, see the firecrawl repo). Any Firecrawl-compatible instance works, local or remote.

Install (local, no npm publishing)

npm pack ./packages/dsh-native-web
npm install -g ./vcxmug-dsh-native-web-0.1.0.tgz

Add a pure mount-point row to your agent preset ($DSH_HOME/.agent-presets//agent.cordis.yml, or edit the preset in the Web UI: Settings → Agent presets):

- id: native-web
  name: '@vcxmug/dsh-native-web'

Configure in the Web UI: Settings → Plugins → dsh-native-web (baseUrl — empty auto-probes 3002 —, apiKey, apiVersion, timeouts). Changes apply on the next tool call — no session restart.

Start a NEW session with that preset. The model sees native_search and native_scrape and prefers them for fresh, local, token-cheap web access.

Why native instead of MCP?

• MCP: session → DSH MCP client → firecrawl-mcp process → instance API.
• Native: session → instance API (one direct call).

No intermediate process, no protocol bridge, no extra tool-lifecycle turns — and no cloud search provider, so fresh results stay on your own instance.

Known limitations

• native_scrape health depends on the instance's playwright service; a misbuilt playwright service can return empty markdown while search still works (see in the repo root).