To add a web-search MCP server to Open WebUI, an admin opens Settings, Admin, Integrations, adds an External Tool Server of type MCP (Streamable HTTP) with the server URL and a Bearer key, and saves. Size the tool list to the model: our 28 tools measured 13,947 tokens on 2 October 2026, the three web tools 4,215.
Open WebUI is the chat interface many people put in front of Ollama, and since v0.6.31 it supports MCP natively, without the mcpo proxy older guides still describe. That makes it the easiest way to give a whole team's local models web access: one admin adds the server, access control decides who gets it, and every chat can switch the tools on. This guide does that with a hosted web-data server and explains the settings that make it work on small models.
Before you start
| Requirement | Why |
|---|---|
| Open WebUI v0.6.31 or newer | Native MCP arrived in that release; Open WebUI's own MCP support is Streamable HTTP only |
| An admin account | MCP servers can only be added by administrators, by design |
| A model with reliable tool calling | Open WebUI's docs: MCP connects the tools, it does not improve the model's ability to use them |
| A QuanticData API key | Sent as the Bearer token; created in the dashboard |
Why admin-only? Open WebUI's documentation explains that an MCP server runs inside its trust boundary with the connecting user's scope, unlike an OpenAPI tool server, so registering one is reserved for admins. Users still get the tools; they just cannot add servers.
The 5 steps
- Open Settings, then Admin, then Integrations.
- Under External Tool Servers, click Add Connection. The Integrations page has several sections; it has to be this one.
- Set Type to MCP (Streamable HTTP), not OpenAPI. Pasting an MCP config into an OpenAPI connection is the documented cause of the infinite loading screen.
- Enter the Server URL
https://api.quanticdata.io/mcp?tools=web, set Auth to Bearer and paste your API key in the Key field. - Save, restarting Open WebUI if it asks.
Then, in a chat, click +, open Integrations, then Tools, and enable the QuanticData tools. To limit who sees them, scope the connection with Access Control to the right users or groups.
Open WebUI also offers an OAuth 2.1 auth mode using dynamic client registration, which our endpoint supports (see how our MCP OAuth works). With OAuth each user connects their own account; Open WebUI warns that OAuth 2.1 tools cannot be set as default tools on a model, and that WEBUI_SECRET_KEY must be set or OAuth tokens break on every container restart. For a single shared key, Bearer is simpler.
Pick the tool set for your model's context
Open WebUI usually sits in front of Ollama, and Ollama gives a model 4k tokens of context on GPUs under 24 GiB of VRAM, 32k between 24 and 48 GiB, and recommends at least 64,000 for web search and agents. Every enabled tool's definition travels with every request. Measured from what our server returns, tokens as characters divided by four:
| Server URL suffix | Tools | Definition tokens | Use with |
|---|---|---|---|
| (none) | all 28 | 13,947 | models at 32K context or more |
| ?tools=research | search, search_and_read, scrape, map, batch, batch_status | 5,122 | 16K and up |
| ?tools=web | search, search_and_read, scrape | 4,215 | 8K and up |
| ?tools=lite | search_and_read | 356 | 4K models |
| ?tools=collectors | list_collectors, run_collector, collector_run_status | 829 | add to any set for ready-made datasets |
Nothing stops you adding the server twice under two names, for example a lite connection for the small default model and a full one for a 64K model, and scoping each with Access Control. Results matter as much: a default search_and_read returned 3,986 tokens, and with top_n 2 and max_tokens 1,500 it returned 1,858. Put those limits in the model's system prompt in Open WebUI so the model sends them. The full budget for 4K to 64K is in local LLM web search with MCP.
Troubleshooting "Failed to connect to MCP server"
- Bearer without a key. Open WebUI sends an empty
Authorization: Bearerheader if the Key field is blank, and servers reject it. Paste the key, or our endpoint answers 401. - Empty Function Name Filter List. Open WebUI's docs list a known bug: if the filter list is empty, try adding a single comma.
- Handshake timeout. The MCP initialize call has a 10-second default timeout (
MCP_INITIALIZE_TIMEOUT). A server with many tools can take longer on a cold start; raise it or use a smaller tool set. - Wrong type. Infinite loading after saving means the connection was created as OpenAPI. Disable it and add it again as MCP.
- The model never calls the tools. Check the tools are enabled in the chat, then the model: switch to one trained for tool calling, and to a smaller tool set if the context is small.
The other local stacks use the same endpoint: giving Ollama internet access for a terminal, LM Studio MCP web search for the desktop app. Every argument of every tool is in the API documentation, and the QuanticData MCP server page lists all clients. Billing is per successful call: every account gets $2 of free API usage each month, and failed requests are never billed.