#Lobster - a composable workflow executor for clawd.

1 messages · Page 1 of 1 (latest)

trim solstice
#

I'm building a worklow runtime for clawdbot. This is so that clawd doesn't have to plan and replan each step of deterministic workflows that are repeatable.

For instance, when you tell clawd triage my email queue, clawd doesn't have to plan the whole execution every time. It can run email.triage (with whatever parameters we provide as ability) and it'll be a single tool call instead of multiple. Something like lobster.run("email.triage")

The new plugin boundary changes that peter made make this possible. I'd like feedback on the direction and the vision.

#

This effectively replaces multiple bash calls, has a json output and piping instead of stdout and just reading bytes.

What limitation does Lobster actually solve?

It's not really "bash is limited" — clawd can already run bash. The real limitations are:

  1. Every workflow requires LLM orchestration
    • Today: user asks → LLM plans steps → calls tools one by one → interprets results → decides next step
    • Each step costs tokens and can go wrong
    • Lobster: deterministic pipeline runs without LLM deciding each step
  2. No standard "ask before acting"
    • Today: you tell clawd "don't send until I confirm" in the prompt (and hope it listens)
    • Lobster: approve is a language primitive that halts execution until resumed
  3. No workflow memory
    • Today: clawd doesn't know what it triaged yesterday unless you build state management yourself
    • Lobster: stateful workflows that track cursors/checkpoints
#

So far I've built the core runtime:

  • Typed JSON pipelines (| pipes objects, not text)
  • Commands: exec, where, pick, head, json, table
  • Approval gates: approve halts, emits token, resume continues
  • Tool mode: single JSON envelope output with protocolVersion
  • clawd.invoke: call clawd tools from Lobster
  • gog.gmail.search, gog.gmail.send: Gmail wrappers
  • email.triage: rule-based classification into buckets
  • doctor, version: health checks

Have a PR out to use the new plugin boundary to integrate with core clawd. Once that's merged lobster can develop independently and add abilities.

The goal is to have clawd use lobster under the hood. Instead of the model figuring out skills every time.

lunar glade
#

i guess this is a "execute on call cronjob + hooks" but formalized and structured as a pipeline? deterministic runs with callbacks... i like the idea

and will probably help a lot when running with cheaper models since you can define the specifics of what needs to happen and won't have to rely on agent intelligence (if i understand correctly)

tawdry frigate
#

this is a startup worthy idea! we're also working on it..

#

leads to composable mcps

#

haven't gotten as far as you conceptually, with my approach I'm looking at tool calls, extracting parameters and essentially replaying them via a higher order mcp

trim solstice
trim solstice
tawdry frigate
#

Let's collab!

trim solstice
#

Sure. Give me a few days to get lobster to a place where I want it to be. My goal is to build the wow path for a few popular plugins. Then it can become a standard way to run persistent stuff on top of clawd.

If our community likes it my plan is to extend this to Claude plugins.

#

I have a PR out to use peters plugin boundary changes. Once that's merged lobster can be developed outside of clawdbot core and independently.

tawdry frigate
#

You're building the next generation of compilers. Excited to see the progress. I'll follow the GitHub

trim solstice
#

Clawdbot integration is pending but I've been hacking more on this and I'm still building it with clawd as the firstclass user. I'm adding TS/JS support so you can essentially drop ten lines of code in and have a deterministic pipeline your AI agent can call.

tawdry frigate
#

Hey @trim solstice , enjoyed your presentation at ClawCon. I think it is difficult to articulate the value proposition of what your are building and you did a good job.

#

What's the best way to help / get involved with the development of Lobster?

trim solstice
#

Thanks. It's also open source. Happy to read issues, feature suggestions, or PRs 🙂

tawdry frigate
trim solstice
#

Yes the idea is to have community workflows similar to skills. I'm intentionally moving slow there while we figure out how to securely run clawdhub

bright stream
#

hey folks. just came here to say this is an absolutely amazing idea! I've got my second brain in notion with a schema that's not exactly simple. at the moment i have a huge .md file explaiing it that openclaw has to read every time and then figure out what to do for a given multi-step workflow. it's very hit or miss.

I just wrote a little cli utility for it to make all this simpler for the LLM: https://github.com/bobbbino/ultimate-brain-cli/

Piping the output of one command to the next is exactly what i need! as it stands the LLM still has to read the output of one and kick off the next.

trim solstice
#

Thank you! I would love to build more into lobster that you think would be useful. Even better if you can build workflows for it. Feel free to send PRs to extend lobster’s abilities!

bright stream
#

I definitely have some ideas and can have a go at coding some of them. How do you want to track FRs? GitHub Issues or just chat here?

trim solstice
#

Issues please so we won't lose track 🙂

bright stream
#

just been doing some reading around this topic. what do you think of anthropic's approach to this of getting agents to write code to chain tool calls together and then saving the working code snippets as new tools in their library?

trim solstice
#

It's in the same direction. Just that it would be nice to have LLM friendly bash. And provide more flexibility for them to design their own repeatable workflows.

unborn void
trim solstice
#

@dusky horizon nice job with executor!