← Back to Discover

openclaw/lobster

An open-source workflow shell for AI agents that chains tools and model calls into typed, local pipelines with approval gates, so an agent can run a whole workflow in one step.

Workflow automationAgent frameworks

What it does

Lobster is a workflow shell built for OpenClaw. It turns skills and tools into pipelines that pass typed data, such as objects and arrays, between steps rather than plain text, and it runs locally. The goal, as the README puts it, is that an agent can invoke a whole workflow in one step to save tokens. Workflows are written as files that mix shell steps, native pipeline steps and approval gates. Steps can pass output forward, run conditionally, retry, time out and loop. Native steps can call a language model through the OpenClaw gateway, Pi or a generic HTTP adapter, or call a configured OpenClaw agent. An approval gate pauses a run until a person decides, and can require a specific approver. It is a TypeScript package with a command-line tool and an SDK, and it needs Node.js 22 or newer. The README says Lobster must not own OAuth or tokens, and it can draw a workflow as a graph in Mermaid, DOT or ASCII.

Capabilities

  • Typed pipelines instead of text pipes
  • Workflow files with shell steps, pipeline steps and approval gates
  • Model calls from workflows, plus calls to configured OpenClaw agents
  • Per-step retry, timeout and error handling
  • Cost tracking with an optional cost limit
  • Workflow graphs in Mermaid, DOT or ASCII

When to use it

  • You run OpenClaw and want repeatable automations an agent can call in one step
  • You need a human approval step between actions in an agent workflow
  • You want local-first pipelines that keep OAuth and tokens out of the workflow tool