exoclaw-tools-llm-call

Single-shot LLM call tool with Jinja2 templating for exoclaw.

No agent loop, no tools — just prompt in, text out. Use with batch for parallel cheap-model processing.

Usage

from exoclaw_tools_llm_call import LLMCallTool

llm_call = LLMCallTool(
    provider=provider,
    allowed_models=["haiku", "sonnet"],
    default_model="haiku",
)
app = Exoclaw(tools=[llm_call, ...])

The agent can then call:

{
  "prompt": "Feed: {{ feed_name }}\n\n{{ file(data_path) }}\n\nExtract interesting URLs as JSON.",
  "vars": {"feed_name": "Simon Willison", "data_path": "/tmp/batch_abc/0.json"},
  "model": "haiku",
  "output": "/tmp/results/0.json"
}

Template features

• {{ var }} — variable substitution from vars dict
• {{ file('/path/to/file') }} — inline file contents
• All Jinja2 features: filters, conditionals, loops

Combined with batch

batch(tool="llm_call", items=[
  {prompt: "...", vars: {data_path: "/tmp/feeds/0.json"}, model: "haiku", output: "/tmp/out/0.json"},
  {prompt: "...", vars: {data_path: "/tmp/feeds/1.json"}, model: "haiku", output: "/tmp/out/1.json"},
  ...
])

93 cheap LLM calls in parallel. Main agent only sees the filtered output.