This guide covers NC’s dedicated Web Search workflow. For generic MCP servers and tools assigned to a persona, use MCP connections and persona tools. Those connections and approvals do not automatically configure Web Search.
Web Search adds current web evidence to the existing main-chat model. It does not replace the chat provider, and it does not install or start its MCP search servers silently. Search is performed by local MCP endpoints, and the retrieved snippets or pages are inserted into the normal chat request.
Setup Versus Activation
Open Addons → Web Search to reach this workspace.
- Addon setup: enable Web Search in the Addons workspace and restart NC if requested.
- Server setup: in Servers, explicitly select Install / repair search servers, start the installed endpoints, and check them.
- Activation: in Overview, enable Enable web search, choose a Conversation search mode, and use the Web Search switch in the conversation header or Controls.
- Verification: run a diagnostic search in Test before relying on a chat answer.
Installing a package is not the same as starting a server. Starting a server is not the same as the conversation switch being on.
Server Requirements
The default provider chain is:
- SearXNG MCP at
http://127.0.0.1:3000/mcp - DuckDuckGo MCP at
http://127.0.0.1:8000/mcp
Both are expected to use Streamable HTTP mode.
SearXNG
The SearXNG MCP package is only a bridge. It requires an operator-provided SearXNG search instance. NC does not install or start SearXNG itself. By default the bridge expects SearXNG at:
http://127.0.0.1:8080
Set SEARXNG_URL for the NC environment if your instance runs elsewhere. Keep
the SearXNG instance and the MCP endpoints loopback-only unless you have
deliberately secured a wider deployment. A SearXNG MCP endpoint that reports
online still cannot search if the underlying SearXNG instance is unavailable.
DuckDuckGo
DuckDuckGo MCP is the no-key fallback. It is self-contained after its addon-local package is installed, but it is still an external service dependency and can be rate-limited or unavailable. The fallback does not make SearXNG unnecessary when you specifically want the SearXNG provider.
Install And Start Servers
Open Servers:
- Review the SearXNG MCP, DuckDuckGo MCP, and Server installation status lines.
- Select Install / repair search servers. This explicitly installs the DuckDuckGo MCP server into an addon-private Python environment and, when Node.js/npm is available, the optional SearXNG MCP bridge into addon-private Node storage.
- Wait for installation to finish. It runs in the background and should not open console windows. It does not modify NC’s main Python environment.
- Select Start local search servers. This starts the addon-local packages.
- Select Check endpoints to probe both configured loopback endpoints without starting anything.
- If needed, use Stop NC-managed servers to stop only processes that this addon started.
An endpoint started outside NC is detected and left under external ownership. NC will not claim or stop it. If you run your own MCP servers, keep their URLs and lifecycle under your control.
Installation can be cancelled by shutting down the addon. A partial install is not a working endpoint; re-run Install / repair and inspect the per-server message.
Test The Chain
Open Test:
- Enter a harmless Search query.
- Select Run test search.
- Read Diagnostic result.
The test runs a backend query without sending a chat turn and shows a provider, title, URL, and snippet, or a concrete error. The test query is sent to the search backend. If the test fails, do not enable chat search and assume it will work later; fix the endpoint or fallback first.
Activate In Chat
Open Overview:
- Enable web search is the global switch. Turning it off immediately blocks
search-intent model calls, normal-chat and Local Room searches, explicit
/webrequests, and diagnostic test searches. It preserves the saved search mode, provider configuration, and running servers. - Conversation search mode selects Off, Automatically when current information is needed, or Every normal chat turn.
- Use LLM for context intent is an optional, provisional automatic-mode
check. It asks the selected chat LLM whether to search and extracts a focused
query, adding one model call per incoming turn when a backend is available.
Explicit
/weband Every normal chat turn bypass that check. - Maximum search results and Pages to read bound retrieval work.
- Read result pages for grounded evidence controls whether result pages are fetched in addition to snippets.
- Allow image searches when requested and Allow video searches when requested are separate retrieval options.
The Conversation search mode and the global switch are separate gates. Both
must allow a search. In particular, mode Off blocks explicit /web queries;
turning the global switch off also blocks them.
Conversation switch
In the Chat header, Web Search (also available from Controls if hidden) is the activation control for the conversation:
- Turning it on starts installed addon-managed servers and enables chat search.
- If no servers are installed, NC offers setup rather than silently installing.
- Turning it off disables chat search and stops NC-managed search servers.
- Externally managed endpoints keep their own lifecycle.
Do not confuse this with the global Enable web search switch. The global switch blocks search but preserves the saved provider setup and running-server state; the conversation switch is the per-conversation activation and can stop NC-managed servers.
Commands
| Command | Effect |
|---|---|
/web status |
Show search state |
/web auto |
Search when current information is needed |
/web always |
Search on every normal chat turn |
/web off |
Disable search mode, including explicit /web searches |
/web clear-cache |
Clear cached web evidence |
/web <question> |
Force a one-turn search |
Commands and the UI switches affect the normal chat path. A forced /web query
does not override the global off switch or conversation mode Off.
Evidence, Cache, And Privacy
Search results and fetched pages are untrusted source data. They can contain instructions, stale claims, or malicious text. The selected model should use them as evidence, not as higher-priority instructions. Verify links, dates, and quotes before relying on an answer.
- Search result instructions is guidance prepended to retrieved evidence. It does not make the sources trustworthy.
- Clear search cache removes stored evidence without changing provider settings.
- A search sends the query to the chosen backend. Fetching pages sends requests to those sites.
- Search provider keys and MCP endpoint URLs are not stored in presets or session data. Do not put credentials in a search query.
- Keep local endpoints loopback-only for a normal desktop setup.
Troubleshooting
| Symptom | Check |
|---|---|
| Search is on but nothing is retrieved | Check the conversation switch, global switch, and mode; run Test |
| SearXNG endpoint is offline | Start the SearXNG instance, verify SEARXNG_URL, then start/check the MCP bridge |
| SearXNG MCP is installed but cannot search | The bridge is running but the underlying SearXNG instance is not available |
| DuckDuckGo fallback fails | Check endpoint readiness, network access, and rate limiting; the fallback is not guaranteed |
| Install / repair is unavailable | Wait for any current installation to finish; re-read the installation message |
| External server was not started or stopped | NC only manages processes it started itself |
/web is rejected |
The global switch or conversation mode may be off |
| Answers cite poor sources | Adjust result and page limits, use a better SearXNG instance, or verify the links yourself |
Current Limits
- SearXNG MCP requires a separately running SearXNG instance. NC does not install, configure, or launch SearXNG.
- The DuckDuckGo fallback is convenient but not a substitute for a reliable private search backend.
- Search introduces latency and can add a model call in automatic mode.
- Automatic mode uses freshness heuristics unless Use LLM for context intent is enabled; neither is a guarantee that every needed fact is found.
- Fetched pages and snippets are untrusted and may be incomplete.
- Search does not make the LLM’s answer authoritative; verify important claims.
- Image and video search permissions do not guarantee a usable result.
