LLM Task Tool for Workflows

llm-task ships as an optional plugin tool that executes a single LLM call limited to JSON, producing structured output that can optionally be validated against a JSON Schema. This lets workflow engines such as Lobster incorporate an LLM step without requiring custom OpenClaw code for each workflow.

Enable

Activate the plugin:

{
  "plugins": {
    "entries": {
      "llm-task": { "enabled": true }
    }
  }
}

Grant access to the tool:

{
  "tools": {
    "alsoAllow": ["llm-task"]
  }
}

alsoAllow extends the active tool profile with llm-task without blocking any other core tools. Only reach for tools.allow when you need a strict allowlist approach.

Config (optional)

{
  "plugins": {
    "entries": {
      "llm-task": {
        "enabled": true,
        "config": {
          "defaultProvider": "openai",
          "defaultModel": "gpt-5.6-sol",
          "defaultAuthProfileId": "main",
          "allowedModels": ["openai/gpt-5.6-sol"],
          "maxTokens": 800,
          "timeoutMs": 30000
        }
      }
    }
  }
}

allowedModels defines which provider/model strings are permitted; any request for a model outside this list gets rejected. Every remaining key acts as a per-call fallback, applied when the tool call leaves that parameter unspecified.

Tool parameters

Parameter	Type	Notes
`prompt`	string	Required. Describes the job the LLM should perform.
`input`	any	Optional data payload; serialized into JSON and appended to the prompt.
`schema`	object	Optional JSON Schema that the parsed output must conform to.
`provider`	string	Replaces `defaultProvider` or the agent's default provider.
`model`	string	Replaces `defaultModel`; accepts plain model IDs, aliases, or a `provider/model` reference (a duplicate provider prefix gets removed automatically).
`thinking`	string	Reasoning level (for instance `low`, `medium`); the resolved model must support the chosen level.
`authProfileId`	string	Replaces `defaultAuthProfileId`.
`temperature`	number	Best-effort; not every provider respects it.
`maxTokens`	number	Best-effort cap on output tokens.
`timeoutMs`	number	Run timeout; defaults to `30000`.

Output

Produces details.json (the parsed JSON that passed schema validation) together with details.provider and details.model, which identify the model and provider that actually executed.

Example: Lobster workflow step

Important limitation

The following example assumes the standalone Lobster CLI is active and that openclaw.invoke already holds the correct gateway URL and authentication context.

For the embedded Lobster runner bundled inside OpenClaw, this nested CLI pattern is not currently reliable:

openclaw.invoke --tool llm-task --action json --args-json '{ ... }'

Until embedded Lobster exposes a supported bridge for this flow, choose either:

direct llm-task tool calls outside Lobster, or
Lobster steps that avoid nested openclaw.invoke calls.

Standalone Lobster CLI example:

openclaw.invoke --tool llm-task --action json --args-json '{
  "prompt": "Given the input email, return intent and draft.",
  "thinking": "low",
  "input": {
    "subject": "Hello",
    "body": "Can you help?"
  },
  "schema": {
    "type": "object",
    "properties": {
      "intent": { "type": "string" },
      "draft": { "type": "string" }
    },
    "required": ["intent", "draft"],
    "additionalProperties": false
  }
}'

Safety notes

JSON-only: the model is instructed to return a plain JSON value, with no code fences or extra text.
No tools: the underlying run has tools turned off, so the model cannot invoke anything mid-task.
Treat output as untrusted unless you validate it with schema.
Place approvals before any side-effecting step (send, post, exec) that uses this output.