Skip to content

Built-in Agent Tools ​

L2M ships a few tools agents often need, served by the API itself. You don't run an MCP server for them: drop a preset on an agent's Tool port and pick a key if the provider needs one.

ToolTools the agent seesKeyCalls go to
Wikipediawikipedia_search, wikipedia_pagenonehttps://<language>.wikipedia.org (default en)
SearXNG web searchsearxng_searchoptional bearer tokenyour SearXNG instance
SerpAPI web searchserpapi_searchSerpAPI API keyhttps://serpapi.com
Wolfram Alphawolfram_alphaWolfram Alpha AppIDhttps://www.wolframalpha.com (LLM API)

Add one ​

  1. Open the node drawer from an agent's Tool port, or the palette.
  2. Pick Wikipedia, SearXNG web search, SerpAPI web search or Wolfram Alpha. Each is an MCP Tool node pointed at the built-in l2m-tools adapter, with one provider set.
  3. For SerpAPI and Wolfram Alpha, choose the key under Auth Secret. For SearXNG, set Instance URL to your instance; its JSON output format must be enabled (search.formats includes json).
  4. For Wikipedia, set Language edition (de, ja, ...) if you don't want English.

The node works on its own too, outside an agent, like any MCP Tool node.

How they are governed ​

The built-in tools go through the same MCP layer as any other server, so nothing about them is special-cased:

  • Tool rules match them with server l2m-tools (inline node) and their tool names.
  • Labels. Every result counts as outside content (untrusted_external). A rule such as Untrusted content: write tools need approval asks before a write tool runs after a search.
  • Read-only. The tools declare readOnlyHint: true and openWorldHint: true. The workflow's write-tool approval does not ask before them.
  • Destinations. Each call is judged against the provider's real host, so data destination rules and content rule destination checks see en.wikipedia.org, serpapi.com or your SearXNG host.
  • Outbound policy. Requests go through the API's outbound URL checks: with OUTBOUND_DENY_PRIVATE=true, a SearXNG instance on a private address is refused unless OUTBOUND_ALLOWLIST names it. Redirects are checked hop by hop.
  • Keys. The key is read from the secret when the call is made, after the checks above, and added to the request by the server. It is never a tool argument, so the model, argument summaries, telemetry rows and recorded cassettes never see it; error messages that quote it are scrubbed.
  • Telemetry, replay. Calls count in MCP tool telemetry, and a recorded run replays them from its cassette without calling the provider. MCP server rate limits and the circuit breaker are set on registered servers, so they do not apply to these preset nodes; watch the provider's own quota.
  • Write-tool approval skips them only because they are L2M's own adapter. An admin's manual read-only setting, or a tool rule on server l2m-tools, still decides.

Answers are trimmed (at most 20 results, long text cut). A response is read as it arrives and stops at 2 MiB, so an answer with no length header cannot grow past that.

Endpoint overrides ​

Instance URL (SearXNG, and the endpoint field of the other providers) sends the calls to the host you set:

  • Destination rules and content rule destination checks judge that configured host.
  • Redirects are followed by the outbound URL checks, which check each hop against OUTBOUND_DENY_PRIVATE and OUTBOUND_ALLOWLIST and drops credentials on a cross-origin hop. Data destination rules judge the configured host, not each redirect hop.
  • The tool still counts as read-only: it sends a GET with the model's query. Point an override only at a host you trust with that query.

One name per tool ​

Each tool an agent sees must have its own name. Two Wikipedia nodes on one agent (two languages, or one with another endpoint) would give the model l2m-tools__wikipedia_search twice, so the agent node fails with an error naming the duplicates, and validation reports duplicate_agent_tool_name. Use one node per provider on an agent, or split the work across two agents. Other MCP servers that serve one tool name are renamed instead; see Tool names the model sees.

Think tool ​

Turn on Think Tool on an Agent or Supervisor to give it a think tool: a scratchpad step for planning and for checking a tool result before acting on it. It is off by default.

  • Calling it fetches nothing and changes nothing. The model gets a short acknowledgement.
  • Each note appears in the node's activity and the run tape as Thought: ..., and on the step's tool results with a thought field.
  • It is a runtime tool like agent_clarify: no tool rules, content rule checks or labels apply to it (the note is the model's own output, and the final answer is still checked), it is not cached in the session tool cache, and it writes no MCP telemetry row.

This website (not the L2M product) uses Google Analytics, only after you accept cookies. Cookie settings