dsh-web-search-anysearch
A DeepSeek Harness (dsh) plugin that
provides web search through AnySearch — "AI Search Infrastructure
for Agents".
Requires dsh >= 0.2.0-rc.2 — this plugin targets the dsh RC/stable line only (CI and releases resolve the newest of the latest/next dist-tags at runtime). The alpha line is no longer supported.
The plugin registers a WebSearchProvider (id: "anysearch") with the official
ctx.web capability seam,
so the model keeps calling the same web_search tool with the same citation formatting
and bounds. No third-party SDK is used: the transport follows AnySearch's own reference
CLI (anysearch-ai/anysearch-skill).
How it talks to AnySearch
Per the official interface spec (scripts/shared/doc_spec.md in the skill repo):
POST https://api.anysearch.com/mcp — JSON-RPC 2.0 tools/call, stateless (no
initialize, no session, one connection per request).Authorization: Bearer <API_KEY> is optional — anonymous access works with lower
rate limits. A free key is available at anysearch.com/console/api-keys.max_results is capped at 10 by the API.- Responses are structured Markdown for agents; the provider parses result blocks into
citeable sources, and degrades to passing the whole payload through as
content if the
format ever drifts — a search never fails just because the parser did.
- Raw
node:https per request (not fetch/undici), matching the official CLI and
avoiding a known undici connection-pool hang on this endpoint.
Vertical routing (domain/sub_domain via get_sub_domains), batch_search, and
extract remain richer MCP/skill-layer features; this provider deliberately exposes the
general search tool to dsh's model-facing web_search (a domain can be pinned in
config). Point an MCP client at AnySearch for the full toolset.
Install
One command:
dsh plugin --profile <name> add @aiwayds/dsh-web-search-anysearch
dsh plugin installs the package into the profile (via pnpm) and registers it in
dsh.profile.bundles automatically — the package declares dsh.bundle, so the
official CLI reconciles the layer stack by itself. Restart dsh. That is the whole
install: the plugin mounts itself AND claims the web_search default at the bundle
layer, so with zero configuration the model's web_search tool runs on AnySearch
(anonymous access needs no key).
Local development: dsh plugin --profile <name> add link:/path/to/dsh-web-search-anysearch
(relative path specs anchor to your invoking directory). The manual route — adding the
dependency and the bundle entry to ~/.dsh/profiles/<name>/package.json yourself and
running pnpm install there — works identically.
@deepseek-ai/* packages are peer dependencies by design: they must resolve to
the profile's single shared dsh closure. Putting them in dependencies installs a
second cordis instance and crashes the loader.
Uninstall
dsh plugin --profile <name> remove @aiwayds/dsh-web-search-anysearch
The host auto-cleans the whole install: the bundles entry is spliced out of the
profile's package.json, the dependency is removed, and the plugin's patch
layer drops with the package — including the searchProvider: anysearch
claim the patch placed on the host web row, so web_search reverts to the
host default (deepseek-official). Restart dsh and it is as if the plugin was
never installed.
There is no on-disk state to clean up: the plugin is stateless (no cache, no
config writes). The optional API key, if you ever stored one, lives in the
host-owned credentials document — remove it there if unwanted.
One caveat: if you hand-copied the selection into your own patch layer
(searchProvider: anysearch in the profile or home cordis.patch.yml), that
line outlives the plugin — searches then fail gracefully with
WEB_PROVIDER_UNAVAILABLE (no provider with that id is registered) until you
delete the line.
Configure (all optional)
API key — only needed for higher rate limits. Anonymous access is valid AnySearch
usage; when the anonymous quota runs out, the search error names the credential to
configure. Store the key in dsh's managed credentials document
(~/.dsh/.credentials.yaml):
version: 1
refs:
ANYSEARCH_API_KEY: <your_key>
Or fill it from the GUI: Settings → Plugins → the dsh-web-search-anysearch bundle →
its web-search-anysearch row → Configure. That page writes the reference the
row's own configuration names — apiKeyEnv, or ANYSEARCH_API_KEY when config
names none — through dsh's credentials service and reports whether a key is
currently configured. Because that field is live, re-pointing apiKeyEnv moves
the page and the provider together instead of splitting them — a settings write
applies to the next search, a patch edit at the next start — and a key staged on
the page is dropped if the reference changes under it. The field is
write-only — a stored key never rides a response back to the page — and a blank
field keeps the current key. One more click
(Clear key, which asks once) removes the reference again, returning searches to the
anonymous limits. A key supplied by the launching environment or a .env fallback is
read-only, and the page says so instead of offering a write.
Both routes land in the same document under the same reference, so they are
interchangeable: a key typed in the GUI is visible to a hand edit and vice versa.
Nothing is written into a configuration file, and the provider re-resolves the
reference on every search, so no restart is needed either way.
Keep another engine as the default? Your own patch layer always outranks the
plugin's bundle patch — one entry, no need to touch the plugin:
# profile cordis.patch.yml or ~/.dsh/cordis.patch.yml
- id: web
config:
searchProvider: tavily # or deepseek-official, exa, ...
Verify which engine owns web_search:
dsh --profile <name> --dump-config | grep -A3 'id: web$'
Plugin config (all optional; via a patch layer, read at startup):
| Field | Default | Notes |
|---|
apiKeyEnv | ANYSEARCH_API_KEY | Credential reference resolved per search. Live: the GUI row reads the accepted value and manages that same reference; a blank value names none, so both halves use the default |
apiKey | — | Literal key; wins over the reference |
endpoint | https://api.anysearch.com/mcp | Env override: ANYSEARCH_ENDPOINT |
timeoutMs | 30000 | Per-request budget |
domain | — (general) | Pin a vertical: finance, academic, code, … (17 total) |
maxResults | 8 | Default when the tool sends no bound; API caps at 10 |
Behavior
- Markdown result blocks map to sources:
title, url, snippet text; relative
date:/sitelinks: annotations stay inside the snippet (no ISO dates → publishedAt
is never set); duplicate URLs dedupe.
- Seam error contract: cancellation as
WEB_ABORTED; transport failures, non-2xx
statuses, JSON-RPC errors, and isError tool results as WEB_PROVIDER_ERROR with the
provider's message surfaced.
available() is a local check only (parseable endpoint) — anonymous access means no
key is required to be usable.
Develop
npm install
npm run check # tsc --noEmit
npm test # build + node --test (transport runs against a local HTTP server;
# markdown fixtures come from real API payloads)
After editing src/, run npm run build and restart dsh.
The browser half (src/client.ts → lib/client.js) is a classic script the Web page
evaluates directly, so npm run build also runs scripts/build-client.mjs: TypeScript
always marks a .ts file in this package as a module, and that step strips the marker it
adds and fails the build if the artifact ever carries module syntax. npm test covers
the Client half too — artifact shape, the row page it claims, and the credential writes
it performs.
License
MIT
