Web search for local models in Open WebUI: Exa or Tavily

Install + handshake last tested 2026-10-22 · Local-model tests: Not yet tested

Client: Open WebUI (native MCP from v0.6.31, Streamable HTTP only). Client versions and features come from vendor docs, checked 2026-10-22. We haven't run this setup end-to-end yet.

Prerequisites

• Open WebUI v0.6.31 or later
• Either the vendor's hosted Streamable HTTP endpoint, or mcpo to expose the stdio server
• An Exa or Tavily account if you exceed keyless/free limits (free-tier limits apply)

Steps

1. Open WebUI only speaks Streamable HTTP for MCP. Use the vendor's hosted endpoint (URL per vendor docs), or wrap the stdio server with mcpo.
2. Add it under Open WebUI's tools/MCP settings.
3. Enable only the suggested search tools.

Config

{
  "mcpServers": {
    "exa": {
      "command": "npx",
      "args": [
        "-y",
        "exa-mcp-server"
      ]
    },
    "tavily": {
      "command": "npx",
      "args": [
        "-y",
        "tavily-mcp"
      ]
    }
  }
}

Suggested tool allowlist

Picked from the real tool names in our test; whether a small model succeeds with them is not yet tested.

• Exa (2 tools listed): web_search_exa, web_fetch_exa
• Tavily (5 tools listed): tavily_search, tavily_extract

Measured context cost

• Exa: ~414 tokens (5% of 8k, 1% of 32k) for all 2 tools. Small-model friendly: medium.
• Tavily: ~1,612 tokens (20% of 8k, 5% of 32k) for all 5 tools. Small-model friendly: high.

Method: tiktoken cl100k_base over compact JSON of name+description+inputSchema (approximate). Cost for the allowlist subset alone: not yet measured.

Success rates on small models

• Exa: Not yet tested
• Tavily: Not yet tested

These were our own task runs, with all tools exposed. The setup on this page (client, model, allowlist) hasn't been run end-to-end.

Known failure modes (community-reported, not measured by us)

• Too many tools: the model picks the wrong one. Fix: use the allowlist above.
• Gemma thinking mode suppresses tool calls. Fix: turn thinking off.
• Context overflow from large tool schemas. Fix: fewer tools, or a bigger context window.

FAQ

Does Exa / Tavily work with local models?

Not yet tested. We have only run our standard install and handshake tests (last tested 2026-10-22); tool-call success with local models has not been measured.

How many tools does it expose?

Exa: 2 tools in our tools/list test; Tavily: 5 tools in our tools/list test.

Why limit the tools?

Community reports say small models choose the wrong tool or emit malformed arguments when shown many tools. We haven't measured this yet.

Are there affiliate links on this page?

No.