openclaw - 💡(How to fix) Fix Tavily plugin tools (tavily_search, tavily_extract) not accessible in agent runtime [4 comments, 1 participants]

Bug Description

The Tavily plugin (@openclaw/tavily-plugin) is loaded and the skill shows as ready, but the tavily_search and tavily_extract tools are not accessible to the agent runtime. When attempting to invoke tavily_search, the agent receives "unknown tool" or similar error.

Environment

• OpenClaw: 2026.3.23 (also affected 2026.3.22)
• Gateway: systemd service, loopback only
• Plugin: stock:tavily/index.js v2026.3.22 (loaded)
• Skill: tavily shows as ready (source: openclaw-extra)
• API key: configured in plugins.entries.tavily.config.webSearch.apiKey

Verification

$ openclaw plugins list | grep tavily
│ Tavily       │ tavily   │ openclaw │ loaded   │ stock:tavily/index.js    │ 2026.3.22 │

$ openclaw skills list | grep tavily
│ ✓ ready     │ 🔍 tavily             │ Tavily web search, content extraction... │ openclaw-extra     │

Expected Behavior

Agent should be able to call tavily_search and tavily_extract tools directly, e.g.:

Use tavily_search to search for coffee pods with query="..." and max_results=3

Actual Behavior

Agent reports: "The tavily_search tool doesn't exist as a callable function in my toolkit" and must fall back to direct API calls via exec/curl.

Code Evidence

The tools ARE registered in the plugin code. In dist/pi-embedded-CswW9luA.js:

• Line 79157: createTavilySearchTool returns { name: "tavily_search", ... }
• Line 79194-79195: api.registerTool(createTavilySearchTool(api)) and api.registerTool(createTavilyExtractTool(api))

The tools exist in the compiled code but are not accessible at runtime.

Workarounds Considered

1. Direct API calls (current working method)

Agent uses exec to call https://api.tavily.com/search directly with curl/python. This works but:

• Not user-friendly
• API key must be extracted from config manually
• Error handling falls on the agent

2. Set Tavily as web_search provider

Adding to openclaw.json:

"tools": {
  "web": {
    "search": {
      "provider": "tavily"
    }
  }
}

This makes web_search use Tavily automatically, but doesn't expose the tavily_search/tavily_extract tools with their advanced parameters (search_depth, topic, include_domains, include_answer, time_range, etc.).

3. Potential Fixes

Option A - Gateway restart did not resolve, so this may be a deeper tool registration/routing issue.

Option B - The skill system may be disconnected from plugin tool registration. The SKILL.md documents the tools but the skill system may not be wiring them to the runtime.

Option C - There may be a tool allowlist/denylist conflict. Current config has tools.deny: [sessions_spawn, sessions_send] which shouldn't affect tavily, but worth investigating.

Option D - The bundled web search provider list (BUNDLED_WEB_SEARCH_PLUGIN_IDS) includes tavily, but there may be separate registration paths for "web search providers" vs "explicit plugin tools."

Impact

• Users cannot access Tavily's advanced search features (search_depth, topic filters, AI answers, time ranges)
• Users cannot use tavily_extract for content extraction from specific URLs
• Falls back to basic web_search or manual API calls

Suggested Investigation

1. Trace tool registration flow: plugin registerTool() → tool catalog → runtime exposure
2. Check if skill "ready" status is based on config only, not actual tool availability
3. Verify the skill's SKILL.md docs match what the plugin actually exposes