dsh-searxng-web

English | 简体中文

DeepSeek Harness plugin that backs the native web_search / web_fetch tools with your own self-hosted SearXNG instance — keyless, private, no third-party search vendor.

model ── web_search ──▶ ctx.web ──▶ searxng-web provider ──▶ your SearXNG ──▶ engines
model ── web_fetch ──▶ ctx.web ──▶ searxng-web-fetch ──▶ target page (SSRF-guarded)

Why

• dsh ships web_search pointed at the DeepSeek cloud search, and mounts no fetch provider at all. If you run SearXNG for privacy or offline use, your queries still leak to a vendor — until this bundle is installed.
• Unlike an MCP server integration, this rides dsh's own provider seam: the model keeps using the short native tool names (web_search, web_fetch), every agent and subagent inherits it, and nothing extra runs alongside dsh.

Requirements

• Node.js ≥ 22
• DeepSeek Harness dsh installed (verified on 0.1.5-rc.2, latest release)
• A reachable SearXNG instance with JSON output enabled (settings.yml → search.formats: [html, json]), verified by:

Documentation

• Installation Guide (简体中文)

• Configuration Guide (简体中文)

• Usage Guide (简体中文)

• Update Guide (简体中文)

• Uninstall Guide (简体中文)

• Changelog

  curl 'http://YOUR_SEARXNG:8080/search?q=test&format=json'

Install

From npm (recommended)

dsh plugin --profile web add dsh-searxng-web

(Replace web with your profile, e.g. tui.) Published from CI with Sigstore provenance; the package ships a prebuilt lib/, so nothing needs to be compiled or allowlisted on install.

or from the repository / a tarball:

dsh plugin --profile web add ./dsh-searxng-web        # source checkout
dsh plugin --profile web add ./dsh-searxng-web-1.0.0.tgz
dsh plugin --profile web add github:maxwell-feng/dsh-searxng-web
# or pin a commit:
dsh plugin --profile web add github:maxwell-feng/dsh-searxng-web#<sha>

Git installs fetch sources: the repository commits the compiled lib/ output, so git installs load without waiting on a registry — the prepare script (npm run build) rebuilds lib/ from source after install. pnpm refuses to run a git dependency's prepare until it is allowlisted; if the first add fails, copy the exact package key pnpm printed into the profile's pnpm-workspace.yaml and re-run the add. See the Installation Guide.

Upgrading

dsh plugin --profile web add dsh-searxng-web@latest
# or from git, to pick up changes before they reach npm:
dsh plugin --profile web add github:maxwell-feng/dsh-searxng-web

0.2.x configurations keep working unchanged — every field added since is optional with identical defaults. Since 0.3.0 configuration is validated at load time (Schemastery schema), a mistyped key fails the boot with an actionable error instead of being silently ignored. 0.4.0 adds the optional baseUrls failover list; single-baseUrl setups are unaffected. 0.5.0 adapts to deepseek-harness 0.1.2-alpha.1 (fiber-scoped provider disposers, post-redirect url reporting) — no config changes required. 0.5.3 adapts to deepseek-harness 0.1.2-alpha.2 — the seam and config rows are unchanged, only the dependency pins move. 0.5.4 adapts to deepseek-harness 0.1.2-alpha.3 — packages/web moved only its version pins in that release, so again no config changes required. 0.5.5 verifies against 0.1.2-alpha.4 (latest master): seam unchanged, no config migration. 0.6.0 adapts to deepseek-harness 0.1.5-alpha.1: seam unchanged, adds a standard prepare build script and the standalone CONFIG / UPDATE / UNINSTALL doc suite — no config changes required. 0.7.0 verifies against deepseek-harness 0.1.5-rc.1: the ctx.web provider seam (packages/web/web/src) is source-identical, vendored @deepseek-ai/cordis 4.0.2 / @deepseek-ai/schemastery 3.18.2 unchanged — no code or config migration required. Node floor raised to >=22 (harness floor is ^22.19). Adds INSTALL / USAGE guides and rewrites CONFIG against the actual schema. 0.8.0 verifies against deepseek-harness 0.1.5-rc.2: adds manifestVersion: 1 under package.json.dsh conforming to @deepseek-ai/dsh-package-manifest and declares "engines.dsh": "^0.1.5-rc.2". 1.0.0 refactors the codebase into pure TypeScript (zero JavaScript tracked), running tests natively via Node --experimental-strip-types.

Installing does three things (via the bundled patch layer):

1. inserts the searxng-web plugin row;
2. points ctx.web at its search/fetch providers;
3. re-enables web_fetch (tool-web.fetch).

Then boot as usual:

dsh --profile web

Usage

New sessions now answer "search xxx" through your instance — no tool-name changes:

• web_search → ctx.web → searxng-web → your SearXNG → configured engines
• web_fetch → ctx.web → searxng-web-fetch → target page (SSRF-guarded, HTML→text)

Verify any time:

dsh --profile web --dump-config | grep -A5 searxng
# or inside a session: call web_search "test" and inspect sources[].url

GUI: Settings → Web Search shows provider readiness.

Pointing at your instance

The default base URL is http://127.0.0.1:8080. Override it (and anything else) in your profile's cordis.patch.yml — the user layer applies after bundle layers:

- id: searxng-web
  config:
    baseUrl: 'http://10.42.1.159:8080'
    timeoutMs: 15000        # per-search budget, ms
    fetchTimeoutMs: 30000   # per-fetch budget, ms
    fetchMaxChars: 200000   # cap on web_fetch output characters
    ssrfGuard: true         # refuse private/loopback fetch targets
    search:                 # forwarded to SearXNG on every query (all optional)
      language: ''          # e.g. 'zh-CN', 'en'
      safesearch: 0         # 0 off, 1 moderate, 2 strict
      # categories: 'general'   # 'news', 'it,science', ...
      # engines: ''             # 'google,bing,ddg', ...
      # timeRange: ''           # 'day' | 'week' | 'month' | 'year'

Patch rows replace config wholesale (no deep merge) — restate keys you want to keep when overriding.

Triple-stack endpoints with automatic failover (0.4.0+)

Home instances often live behind several doors at once — a public IPv4, a public IPv6 and a LAN address. baseUrls takes an ordered list and fails over automatically:

- id: searxng-web
  config:
    baseUrls:
      - 'http://203.0.113.10:8081/s/<KEY>'      # public IPv4
      - 'http://[2409:8a55:…]:8081/s/<KEY>'     # public IPv6
      - 'http://192.168.10.144:8081/s/<KEY>'    # LAN (same door, same key)
    timeoutMs: 15000

Semantics:

• Sticky: attempts always start at the last endpoint that succeeded, so a healthy door is never re-probed after an earlier entry had a hiccup.
• Fail fast on the wire only: connection refused / unreachable / timeout / DNS failure advance to the next endpoint. Any HTTP answer (200, 403, 502…) proves that door is alive and its status is surfaced as-is — no silent masking of auth problems.
• One full pass over the list per call; if every endpoint is unreachable you get a single network error.
• baseUrls wins when both it and baseUrl are set; baseUrl alone keeps working exactly as before.

Configuration reference

API keys & authenticated reverse proxies

Three supported ways to put a gate in front of the instance. Credentials configured here ride only on requests to your SearXNG instance; web_fetch targets (model-chosen third-party pages) always stay credential-free.

1. Header gate (recommended for API consumers):

   config:
     baseUrl: 'http://searx.internal:8080'
     headers:
       X-API-Key: 'your-key'

   Pair it with a caddy check, e.g. caddy-l4/forward_auth or a small middleware that compares the header.

2. Basic-auth reverse proxy (caddy basic_auth, nginx auth_basic):

   config:
     baseUrl: 'http://searx.internal:8080'
     basicAuth:
       username: 'searxng'
       password: 'hunter2'

   Setting both basicAuth and a user-supplied headers.Authorization fails at load time with an actionable error.

3. Path-prefix key (no plugin config needed): if your reverse proxy strips a secret prefix before proxying, just include it in baseUrl, e.g. baseUrl: 'http://host:8081/s/<KEY>'. Works because the search adapter appends /search?... to whatever base you give it.

Node's fetch refuses URLs that embed credentials (http://user:pass@…), which is why auth lives in dedicated config fields instead of baseUrl.

Behavior notes & limits

• Search: maps SearXNG results to {url, title?, snippet?, publishedAt?} and surfaces the SearXNG answer line when present.
• Fetch: GET with a browser-ish User-Agent; HTML is reduced to readable text; output capped at fetchMaxChars (truncated flag set).
• SSRF guard: validates the initial target only — redirects are followed without re-validation (v1 limitation). The guard also blocks non-http(s) protocols and unresolvable hosts. Disable only on closed deployments.
• Proxy: uses Node's global fetch, which ignores system proxies and proxy env vars by default — SearXNG traffic always goes direct. Provider APIs behind a proxy are unaffected because this plugin talks only to your instance and fetched pages.
• 403 from SearXNG: JSON output is disabled on the instance — see Requirements above.

Migrating from an MCP-based SearXNG integration

If you previously wired SearXNG through an MCP server (e.g. mcp-searxng via a dsh-mcp-client row), remove that integration when you install this plugin:

• The model would otherwise see two overlapping search tools (native web_search and mcp__searxng__searxng_web_search) plus several extra schemas — ambiguous tool selection and ~1–2k tokens of per-request overhead for no search-quality gain (both hit the same instance).
• To remove: delete the dsh-mcp-client insert row from your profile's cordis.patch.yml (HMR unregisters the tools immediately) and optionally npm uninstall -g mcp-searxng.

What you give up: the MCP reader's PDF extraction and section filtering. The native web_fetch covers plain HTML/text pages; if you later need PDF reading again, re-adding the MCP row takes minutes.

Uninstall

dsh plugin --profile web remove dsh-searxng-web

Removes both the dependency and the bundle layer. ctx.web falls back to the base composition (DeepSeek search, no fetch provider).

Development

The plugin is written in TypeScript (src/index.ts); the compiled lib/index.js is committed so installs never need a build.

npm install          # dev dependencies (typescript, @types/node, cordis types)
npm run build        # compile src/ → lib/
npm test             # build + self-contained offline test suite (mock SearXNG)

Release process (maintainers)

Bump version in package.json, add a CHANGELOG.md entry, then:

git commit -am "release: vX.Y.Z"
git tag vX.Y.Z
git push --follow-tags

GitHub Actions runs the standalone test suite and publishes to npm via OIDC trusted publishing (Sigstore provenance) — the same pipeline as dsh-windows-ocr.

License

MIT