@maxwell-feng/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 ≥ 20
-
DeepSeek Harness dsh installed (verified on 0.1.1-rc.2)
-
A reachable SearXNG instance with JSON output enabled
(settings.yml → search.formats: [html, json]), verified by:
curl 'http://YOUR_SEARXNG:8080/search?q=test&format=json'
Install
From npm (recommended)
dsh plugin --profile web add @maxwell-feng/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.
From GitHub
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>
The repository commits the compiled lib/ output, so git installs also load
without any build step or pnpm allowBuilds allowlist.
Upgrading
dsh plugin --profile web add @maxwell-feng/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.
Installing does three things (via the bundled patch layer):
- inserts the
searxng-web plugin row;
- points
ctx.web at its search/fetch providers;
- re-enables
web_fetch (tool-web.fetch).
Then boot as usual:
dsh --profile web
New sessions now answer "search xxx" through your instance. Verify in the
compose output any time:
dsh --profile web --dump-config | grep -A5 searxng
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
| Key | Default | Description |
|---|
baseUrl | http://127.0.0.1:8080 | SearXNG instance URL |
baseUrls | (unset) | Ordered endpoint list with sticky automatic failover (0.4.0+); takes precedence over baseUrl when non-empty — see "Triple-stack endpoints" above |
timeoutMs | 15000 | Per-search attempt budget (ms) |
fetchTimeoutMs | 30000 | Per-fetch attempt budget (ms) |
fetchMaxChars | 200000 | Max characters returned by web_fetch |
ssrfGuard | true | Refuse private/loopback/link-local/CGNAT fetch targets |
search.language | (unset) | SearXNG language param |
search.safesearch | 0 | SearXNG safesearch param |
search.categories | (unset) | SearXNG categories param |
search.engines | (unset) | SearXNG engines param |
search.timeRange | (unset) | SearXNG time_range param |
headers | (unset) | Extra HTTP headers attached to SearXNG requests only (e.g. X-API-Key gates) — never sent to web_fetch targets |
basicAuth.username / basicAuth.password | (unset) | Basic-auth credentials for instances behind an authenticating reverse proxy (caddy basic_auth, nginx auth_basic) |
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.
-
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.
-
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.
-
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 @maxwell-feng/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
@maxwell-feng/dsh-windows-ocr.
License
MIT
