Exa MCP endpoint (mcp.exa.ai) timing out while REST API works

[chindris-mihai-alexandru]

Problem

The Exa MCP remote endpoint (https://mcp.exa.ai/mcp) experiences frequent timeouts and 520 Cloudflare errors when used with OpenCode, while the REST API (api.exa.ai) works without issues.

Error Messages

Timeout Error:

Request timeout after 5000ms

520 Cloudflare Error:

520: Web server returns an unknown error

Configuration Attempted

"exa": {
  "type": "remote",
  "url": "https://mcp.exa.ai/mcp?exaApiKey=...&tools=web_search_exa,deep_search_exa,get_code_context_exa"
}

Root Causes

After investigation, there appear to be two separate issues:

1. Timeout Issue (OpenCode-related)

OpenCode v1.1.1 introduced a 5000ms default timeout. The Exa MCP endpoint can take longer than 5 seconds to respond, causing timeout errors.

Fix: Add "timeout": 31000 to the config

2. 520 Cloudflare Error (Exa infrastructure issue)

The 520 error is a Cloudflare error meaning "web server returns an unknown error" - this occurs when Exa's origin server (behind Cloudflare) returns an empty or unexpected response.

Reference: https://developers.cloudflare.com/support/troubleshooting/http-status-codes/cloudflare-5xx-errors/error-520/

Status: Exa team investigated and resolved this issue on January 5, 2026.

Workarounds

Option 1: Add timeout (for timeout issues)

"exa": {
  "type": "remote",
  "timeout": 31000,
  "url": "https://mcp.exa.ai/mcp?exaApiKey=..."
}

Option 2: Use local MCP server (recommended for reliability)

"exa": {
  "type": "local",
  "command": ["npx", "-y", "exa-mcp-server"],
  "timeout": 60000,
  "environment": {
    "EXA_API_KEY": "your-api-key"
  }
}

The local server uses the REST API (api.exa.ai) directly, which is more stable than the remote MCP endpoint.

Related Issues

• HTTP SSE Endpoint returns 405 Method Not Allowed for standard GET connections (Breaks Generic MCP Clients) exa-labs/exa-mcp-server#86 - HTTP SSE Endpoint issues with remote MCP
• websearch_exa MCP timeout after opencode reinstall code-yeongyu/oh-my-opencode#493 - websearch_exa timeout

Status

• Timeout issue: Resolved with config changes (add timeout parameter)
• 520 Cloudflare errors: Fixed by Exa on January 5, 2026
• Recommended approach: Use local MCP server for best reliability

[kevoconnell]

I can replicate the same issue

[cravenceiling]

I might be an opencode "issue". In v1.1.1 was introduced a default timeout of 5000 ms, so you should try adding the timeout property to your config like this:

"exa": {
  "type": "remote",
  "timeout": 31000,
  "url": "https://mcp.exa.ai/mcp?exaApiKey=...&tools=web_search_exa,deep_search_exa,get_code_context_exa"
}

[chindris-mihai-alexandru]

Update after investigation:

There appear to be two separate issues here:

1. Timeout Issue (OpenCode-related)

OpenCode v1.1.1 introduced a 5000ms default timeout. Adding "timeout": 31000 to the config resolves timeout errors.

2. 520 Cloudflare Error (Exa-side)

The 520 error is a Cloudflare error meaning "web server returns an unknown error" - this occurs when Exa's origin server (behind Cloudflare) returns an empty or unexpected response. This is an Exa infrastructure issue, not OpenCode.

Reference: https://developers.cloudflare.com/support/troubleshooting/http-status-codes/cloudflare-5xx-errors/error-520/

Workarounds

1. Add timeout (for timeout issues):

"exa": {
  "type": "remote",
  "timeout": 31000,
  "url": "https://mcp.exa.ai/mcp?exaApiKey=..."
}

2. Use local MCP server (for 520 errors):

"exa": {
  "type": "local",
  "command": ["npx", "-y", "exa-mcp-server"],
  "timeout": 60000,
  "environment": {
    "EXA_API_KEY": "your-api-key"
  }
}

The local server uses the REST API (api.exa.ai) directly, which is working fine.

Related Issues

• HTTP SSE Endpoint returns 405 Method Not Allowed for standard GET connections (Breaks Generic MCP Clients) exa-labs/exa-mcp-server#86 - HTTP SSE Endpoint issues with remote MCP
• websearch_exa MCP timeout after opencode reinstall code-yeongyu/oh-my-opencode#493 - websearch_exa timeout

[miguelmanlyx]

@chindris-mihai-alexandru

Timeouts getting surfaced as rate limits can make post-incident analysis harder, since the remediation paths are very different.

This usually only becomes obvious if there’s a request-level execution record that distinguishes timeout vs provider rejection; I’ve seen that discussed around gateways like AI Badgr.

[i-m-sid]

There is issue with Exa server, nothing to do with OpenCode

[connorads]

Can confirm that using the local exa MCP server instead of remote has fixed the issue for now
https://docs.exa.ai/reference/exa-mcp#configuring-claude-desktop

    "exa": {
      "type": "local",
      "command": ["npx", "-y", "exa-mcp-server"]
    },

[JakubHojsan]

We are investigating the Cloudflare issue on Exa's side

[JakubHojsan]

It looks like the 520 might be resolved. Feel free to tag me in any Exa issue

[JakubHojsan]

[samyons]

Resolved indeed!

[chindris-mihai-alexandru]

Works again for me too!