Lobster for OpenClaw: Typed Workflows That Cut Agent Token Costs

By Beau JohnsonĀ·Ā·8 min read

Quick answer

Lobster is a workflow shell from the OpenClaw team. You define a multi-step workflow once with shell steps, LLM steps, and approval gates, and your OpenClaw agent runs it in a single command that returns typed JSON. It doesn't have to re-plan every time, so it burns fewer tokens.

In this video I go through Lobster, a GitHub repo built by the OpenClaw team. It's a workflow shell that lets your agent run multi-step tasks without re-planning them from scratch every time. If you run OpenClaw agents every day and want them cheaper, more consistent, and safer, this one's for you.

I recorded this in April 2026. Tools like this change fast, so check the repo's current docs before you build on anything here.

What I cover in this video

  • 0:00: What Lobster is and why it matters that the OpenClaw team built it
  • 0:47: Why agents waste tokens re-planning multi-step tasks
  • 2:19: The GitHub pull request monitor example and its typed JSON output
  • 4:35: How workflow files are structured: run steps, pipeline steps, and approval gates
  • 6:05: The Jacket Advice example workflow and a seven-step deployment idea
  • 7:38: Token savings and what my own setup costs per day
  • 9:09: Calling LLMs inside workflows and referencing earlier steps by ID
  • 12:56: Why it's official, and why it doesn't manage its own credentials

The problem: agents re-plan every multi-step task

When you give your OpenClaw agent a multi-step task, it plans every step from scratch, and it does that every single time. Say you ask it to check a pull request: pull the diff, summarize the changes, and tell you whether to merge. The agent has to work out which tools to use, which API calls to make, what order to make them in, and how to format the answer. All that thinking costs tokens. It can also get messy, and mistakes waste even more money.

Lobster is a typed pipeline engine. You define the workflow once, and your agent runs it in one step. No re-planning and no guessing. In the PR example you just tell the agent to run the GitHub PR monitor workflow and give it the repo and the pull request number. It comes back with clean JSON: what changed, who authored it, whether it's mergeable, the review decision, and what state it's in.

That's the key part. Most agents work with walls of unstructured text and then have to dig the answer out of it. Lobster passes typed objects and arrays between steps, so you know what shape the data is in at every point, and you can pipe it straight into the next step.

How Lobster workflow files work

The workflow files read like scripts, which is great if you're just getting started. Each step has an ID, and there are three main kinds of steps:

  • Run steps call shell commands: pull some data, run a script, call an API.
  • Pipeline steps call LLMs directly inside the workflow. That lets you mix deterministic shell commands with AI reasoning in one flow.
  • Approval gates are a hard stop. The workflow pauses and waits for you to say yes before it continues.

I care a lot about approval gates. Plenty of people are nervous about letting an agent delete files, push code, or send emails without seeing it first. A gate fixes that. I still approve everything in my own setup.

One example in the repo is a workflow called Jacket Advice. It takes a location argument that defaults to Phoenix. Step one fetches the weather as JSON. Step two is an approval gate asking whether you want jacket advice from the LLM. Step three, if you approved, pipes the weather data into an LLM with a prompt asking whether you should wear a jacket, be concise, and return JSON. The whole thing is about 10 lines of configuration.

Now scale that up to something like a deployment workflow. Pull the latest code, run the test suite, hit an approval gate to review the results, deploy to staging, run smoke tests, hit a second approval gate, then deploy to production. If the smoke tests fail, the whole thing stops, so broken code never reaches production.

LLM steps, step references, and credentials

LLM steps are what make Lobster more than a fancy bash script wrapper. You call a model through LLM invoke and pick your provider: OpenClaw, Pi, or any HTTP endpoint. You're not locked into one model or service. One workflow could pull data from an API with a shell command, process it with jq, send it to Claude for analysis, and output structured JSON at the end.

The stdin piping between steps is well thought out too. Any step can reference the output of any earlier step by its ID. So step three can take step one's JSON as input; it doesn't have to go through step two. That makes complex workflows with branching logic much easier to manage.

I also like how it handles security. Lobster doesn't own any authentication, manage OAuth tokens, or have its own credential system. It uses the tools and credentials already in your environment variables. If your agent already has GitHub access, your Lobster workflows use those same credentials, and you don't get a new attack surface.

Why this matters for token costs

My system typically costs about $5–7 a day on average. That's a mix of Anthropic models plus GPT-5.4 running the cron jobs, with 13 main AI agents running 24/7 on a Mac mini. Every multi-step task burns tokens on planning alone. With Lobster the planning is already done, so the agent just runs the steps. If I could save even 20–30% of token costs on my complex workflows, that would add up to real money. If you use Claude or Opus through a subscription on OpenClaw, it could also stretch your usage limits a lot further.

To be upfront: I hadn't installed Lobster on my own setup when I recorded this. I'd only just found the repo. My first idea is my Riza agent, which plans the same sequence every morning: research a topic, write a script, format it, save it to the database. As a Lobster workflow, those steps would already be defined with the right tools and parameters. That should mean fewer tokens and more consistent output. When I recorded, the repo already had about 1,000 stars. Since it's from the official OpenClaw organization, I expect typed workflows to become a core part of using OpenClaw.

Key takeaways

  • Lobster lets you define a multi-step workflow once so your OpenClaw agent runs it in one command instead of re-planning it.
  • Steps pass typed JSON objects and arrays, not unstructured text, so outputs are predictable.
  • Workflows mix shell run steps, LLM pipeline steps, and approval gates that pause until you say yes.
  • Any step can pull from any earlier step by its ID, and LLM steps can use OpenClaw, Pi, or any HTTP endpoint.
  • Lobster reuses your existing credentials instead of managing its own, so it doesn't add a new attack surface.

FAQ

What is Lobster for OpenClaw?

Lobster is a workflow shell and typed pipeline engine built by the OpenClaw team. You define a multi-step workflow once, and your OpenClaw agent runs it in a single step that returns structured JSON.

How does Lobster reduce OpenClaw token costs?

Normally an agent spends tokens planning which tools to call and in what order every time it runs a multi-step task. With Lobster that planning is already written into the workflow file, so the agent just runs the steps.

What are approval gates in Lobster?

An approval gate is a hard stop in a workflow. The agent finishes the earlier steps, then waits for you to approve before it does anything else, like deploying code or sending something.

Does Lobster need its own API keys or logins?

No. Lobster doesn't manage authentication or OAuth tokens. It uses the tools and credentials you already have in your environment variables, so if your agent can reach GitHub, your workflows can too.

Transcript

Read the full transcript

0:00I'm going to tell you about a GitHub repo that is going to change the way that you use Open Claw. And I love doing these videos because it helps people hit the ground running when they get set up with Open Claw, automating AI agents, and how to do it properly. So, I have the GitHub repo linked down below. And this one's different from just all of the other repositories that we cover on the channel. Um this one was actually built by the Open Claw team themselves, so you know it's the real deal. The repo is called Lobster. And yes, it has a little lobster emoji and everything, which is pretty cool. But what the lobster actually does is way more interesting than just the name. It's basically a workflow shell for Open Claw. And if you're sitting there going, "What does this even mean?" I'm going to break it down. So, when you give your Open Claw in a your agent a task,

0:47especially something with multiple steps, it has to plan every single step from scratch, and that has to happen every single time. So, you ask it to check a pull request, for example, or pull the diff, or summarize the changes, and tell you basically if you should merge it or not, your agent has to figure out the order of operations like every single time. So, it has to think through what the tools in how to use them. What I like to call first and what I like to call second. And all of that thinking takes up tokens, and it can get messy. So, that's real money down the drain, especially if it makes mistakes. So, what Lobster does is it fixes that. It's really simple. It's a typed pipeline engine that lets you define the workflows once, and then your agent runs

1:34them in one single step. So, there's no replanning, there's no wasted tokens, and there's no guessing. You basically define it once, and then it works. So, let me give you like a real example, because I think this will click once you kind of get the bigger picture. Let's say you want to monitor a GitHub pull request, right? Without Lobster, you would have to tell your agent, "Hey, go check pull request 1,152 on my repo. Look at the status, tell me if anything changed, check the reviews, and check if it's mergeable. And then your agent has to figure out all of that from scratch. It has to reason about what API calls to make, what order to make them in, and how to format that response. That has to happen every single time. But with this Lobster GitHub repo, you have a workflow

2:19file that says run the GitHub PR monitor workflow. Here's the repo, here's the pull request number, and then it's done in one command. So, it comes back with a clean JSON response telling you exactly what changed and who authored it, and is it mergeable, what is the review decision, what state is it in, everything. Typed data, not some messy text blob that the AI has to parse, like actual structured JSON that you can pipe directly into the next step. That's the key here. Keeping it simple, no over-engineering, and I love this for that fact. People spend so much money just fiddling around and messing with Open Claw and getting it set up correctly. We are literally pouring money down the drain, especially if

3:04you're using Anthropic models. One of the biggest problems I see with AI agents right now is they are working with text all the time. Everything is strings of text. Everything is unstructured. Your agent asks for information and then it gets back a wall of text. Then it has to figure out what part of the text is the actual answer. Lobster works with typed objects and arrays. So, when data flows from one step to the next, it is clean and it is predictable. It's not going to hallucinate some random format. You know exactly what shape the data is in at every step of the workflow. Real quick, before we keep going, if you're watching this and you want to actually build with some of these tools, then I'd just watch videos about them, then you're going to want to check out our community down below, Shipping School. We have a full

3:50Claude code course, a full OpenCloud Course, and four live bootcamps every single week where we actually help you get set up from scratch. Like actually set this thing up, not just watch a tutorial and figure it out by yourself. And we also provide one-on-one coaching so you could book a call with me, we could share screens, and I can help you get Cloud Code or OpenCloud running on your machine. That's it, no fluff. I built this community because watching YouTube's only gets you so far. We launched it just 3 days ago and we have over 55 members. You need people around you who are actually building, people who hold you accountable, and coaches who can help you when you get stuck. I'll put the link in the description down below. Get in now before the price goes up. Here is where it gets more

4:35interesting. As the Lobster workflow files look like scripts, and they are actually readable, which I love for someone just getting started with this. You want to be able to understand this at an easy learning level, right? We don't want to overcomplicate. So, you have steps, each step has an ID, you can use the run commands to call the shell tools, you can use pipeline commands to call the LLMs, and you can add approval gates. So, let me walk you through each one of these because they all solve a different problem. The run steps are straightforward. You just, you know, you're just calling a shell command, you know, you pull some data, you run a script, call an API, the basic standard stuff. But the pipeline steps are where it really gets more interesting because

5:20you can call the LLMs directly inside your workflow. So, you're not just running scripts, you are mixing in deterministic shell commands with AI reasoning in the same workflow. That is pretty powerful. And then there are approval gates, and let me just tell you why approval gates matter so much because right now a lot of people are scared to let their AI agents do real work, am I right? I mean, come on. Getting this thing all set up, you really don't want to have full access to your whole personal life, right? I do get that and I have been there as well. You know, you don't want your agent deleting files or, you know, let's say pushing code or sending emails without you at least seeing it first. Lobster lets you put a hard stop in the middle of any of those workflows. The agent

6:05does, you know, step one, step two, then it hits this approval gate. So, it literally stops and just waits for you to say yes before it keeps going. That's huge for trust and just getting started with Open Claw, that is a nice way to go about it. That's how you go from being nervous about your AI agent actually letting it do real work. And I still do this today in my whole setup. I approve everything. So, I pulled up another example, you know, of their workflows and it is so clean. You have a workflow called Jacket Advice. It takes a location argument, defaults to Phoenix. Step one fetches the weather as JSON. Step two is an approval gate that asks, "Do you want the jacket advice from the LLM?" Step three is if you approved, pipes the weather data into an LLM call

6:52and then it asks, "Given this weather data, should I wear a jacket?" You know, be concise and return JSON. And the LLM responds with the structured JSON file. So, the whole thing is like 10 lines of configuration and that's it. Now, imagine scaling that up to something actually real and practical in your flow. You know, imagine a deployment workflow or, you know, it's step one basically it pulls the latest code from your repo. Step two, it tests the suite. Step three is an approval gate where you review the test results. Step four, it deploys to staging. Step five, it runs smoke tests against staging. Step six is another approval gate where you verify everything that it looks good. And then step seven, it deploys to production. All defined in a simple workflow file. All running through Open Claw in one

7:38single command. And if step five fails, the whole thing stops. You do not accidentally deploy the broken code to production, which can be nice. That is the kind of safety net that makes, you know, this actually trust your setup. I want to talk about the token savings cuz this is a huge issue and it's a big deal for anyone running open claw agents every day like I am. My system, people ask me all the time, typically costs about five to seven dollars a day on average. That's using a little bit of Anthropic models with GPT 5.4 handling of the cron jobs. So, basically every time your agent has to plan a multi-step task, it's burning tokens, you know, just on the planning part. It's burning tokens on the reasoning. It's sitting there going, "Okay, first I need to call

8:23this tool, then I need to parse this output, then I need to call the other tool." With Lobster, all of that planning is already done. So, your agent just executes it step by step. There's no wasted tokens in the reasoning part. You know, I run agents all the time, 24/7 on this Mac mini. I have 13 main agents doing everything from writing video scripts to posting on X, generating thumbnails, and the token costs do add up, but for what I'm getting, I think it's worth it. You know, I'm not going to lie, this is just straight up what it costs. And for what you get, I think it's a steal. So, anything that cuts those costs down though on unnecessary planning is a huge savings over time. You know, that could be thousands of dollars a year. So, if I can even save 20 to 30% of token costs

9:09for most of my complex workflows, that adds up to real money over a month. Now, I'm going to talk about the LLM integration because this is where the Lobster separates itself from just being a fancy bash script wrapper. You could call LLMs directly from inside a pipeline step using the LLM invoke. So, you can choose your provider, you can use open claw as the provider, or you can use Pi, you can use any HTTP endpoint. So, you're not locked into one model or one service. You basically can have a workflow that pulls data from an API using a shell command, and then it processes it with jq or some other tool, and then it pipes it into Claude for analysis, and then it outputs a structured JSON at the end.

9:54All in one workflow file. All callable from within Open Claude in one step. You mix and match your tools, and your AI does the rest. That's the vision. Here is where I think something, you know, a lot of people miss about this, you know, when they first look at it, especially if you're getting into this space fresh out the kitchen, you don't know what you're doing. The stdin piping between steps is really well thought out, and if you don't know what that means, it's okay. Each step can reference the output of any previous step by its ID. So, step three can say, "Give me the JSON output from step one as my input." It is not just linear where step two has to use step one, and step three has to use step two. You could just reference it by its step by

10:39the ID tag. So, that makes complex workflows with like branching logic actually more manageable. You're not passing data through a chain, you are pulling it from wherever you need it. If you're struggling to keep up with content, well, I'm about to save you about 40 days worth of work. >> [music] >> I built something called Content Machine. It's 10 AI agents that run on the Open Claude orchestration, and they handle everything. Scripts, thumbnails, X posts, blogs, outreach, clips, newsletters, all of it. So, I went from 1,000 subscribers to 4,000 subscribers on YouTube in 7 days using this exact system. Every single morning, I wake up, and the content's already done. I spend maybe 15, 20 minutes reviewing and

11:25approving them, and I move on with my day. It works [music] for any niche, fitness, finance, real estate, marketing, whatever you are building, and it is 100% completely customizable to your use case. So, you get the mission control dashboard, all of the cron jobs, everything I've built over the last 40 days helping me gain more and more people to subscribe [music] and join the community. So, you plug in your own thing, and it molds it to you, it learns how you talk, and it writes [music] so it doesn't sound like AI slop. $97 one-time, it's not a subscription. I'll put the link down below, and you'll thank me later. I should be straight up with you about this. I have not installed Lobster on my own setup yet. I just found this repo today, but I am 100% going to be trying

12:11it out this week, as well as Meta Claw. We talked about that yesterday, because the use cases for the content pipeline alone are pretty massive, if you think about it, right? Like, that could save you tons of money down the road. My Riza agent has to plan how to research a topic, write a script, format it, and then save it to the database every single morning. What if I could define that entire sequence as a Lobster workflow? The research step, the writing step, the formatting step, the database save step, all predefined with the right tools in the right parameters. Riza just runs the workflow. Way fewer tokens, way more consistent output, way less chance of something going wrong. And if you're This is kind of a touchy subject, but if you're using Claude or Opus uh through

12:56the subscription on Open Claw, this could dramatically cut your usage limits. So, the repo already has 1,000 stars on GitHub, and it was just published this week. And the fact that it is coming from the official Open Claw organization tells me that this is going to be a first-class feature going forward. This is not some random community project that might get abandoned in 2 months. This is from the Open Claw team, and they're saying, "Hey, you know, we think we typed workflows should be a core part of of agent experience." And I agree. I want to point something out that I think is really smart about how they built this. Lobster does not own any authentication. It does not manage OAuth tokens. It doesn't have its own credential system. It uses whatever tools and credentials you have already set up in your

13:42environment variables. So, if your Open Claw agent already has access to GitHub, Lobster workflows can use GitHub through the same credentials. It doesn't need to create a new attack surface. And after security stories we have covered this on the channel, you know how much I care about that. That's the right way to build something like this. And for anyone in the community, Shipping School, this is the kind of thing that we dig into every week. How to make your agents more efficient, more reliable, more cost-effective. And we have 180 members that have joined in the last 18 days. So, we're shipping real products, we're helping each other. You know, if you're serious about building with AI and you need a community of people to hold you accountable, I'll put the link in the description down below. And now we do eight boot camp calls every single week. We're trying to get up to about 20

14:27boot camp calls every week because we want to hit everybody in every time zone so everyone can get value, jump into one of these calls, and get their problems solved. So, that's the vision for the group, and we appreciate you being there. And if you're not building with AI, what are you waiting for? I mean, come on. Subscribe to the channel. We release videos like this four to six times a day, and we'll see you in the next one. Have a blessed day. Who is that sending a message?

Ready to start building with AI?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

Keep reading