Lobster pipelines: deterministic AI workflows

What is Lobster?

Lobster is OpenClaw's pipeline tool. It lets you define multi-step workflows as .lobster files — deterministic sequences of steps where each step receives the output of the last.

Why not just prompt the LLM?

Because "do steps 1 through 8 and check each one" is asking an LLM to hold working memory across a long chain. It drifts. Lobster doesn't drift.

A simple example

pipeline: weekly-report
steps:
  - name: fetch-data
    tool: google-sheets
    action: read
    sheet: "Weekly KPIs"

  - name: summarise
    tool: llm-task
    prompt: "Summarise this week's KPIs in 3 bullet points: {{fetch-data.output}}"

  - name: send-report
    tool: telegram
    message: "{{summarise.output}}"
    chat: "@team-channel"

Key concepts

• skipped: true — any step can emit this to short-circuit the rest cleanly.
• llm-task — the plugin for calling the LLM from inside a pipeline step.
• Step outputs — referenced with {{step-name.output}} in later steps.

When to use Lobster vs free-form prompting

Getting started

Create pipelines/my-first.lobster in your workspace, then trigger it:

openclaw pipeline run my-first