CodaClaw vs OpenClaw vs Claude Cowork

1

Oscar Mike Golf! (O.M.G)

So many of my clients are asking me to help them get into :lobster:OpenClaw - it’s crazy!

And I refuse to assist. Watch my lips… :warning: IT IS NOT SAFE !!! :warning:

Especially if you are not a seasoned developer with a deep understanding of web security.

Instead, I direct them to Claude Cowork (or Claude Code if you know any programming).

It is almost as powerful, is way safer (especially for the naive and bewildered), but lets you learn all the skills needed to keep up to date with AI developments.

And the more I develop real-world projects using Claude Code, the more I see the patterns that are useful for business users.

So I am pivoting my " Coda Vibing Pack" (see here) into a "CodaClaw" pack that works just like Claude Coworker but
(a) keeps all the SKILLS, REFERENCES, SCRIPTS, and ASSETS as Coda Doc Pages, and
(b) uses the Superhuman Go Pack SDK to ‘execute’ all the AI stuff.

It’s very like my previous experience with making AI agents 2 years ago. At first, I built them using “LangChain” and Python (see my experiences here and the technique here). I saw the recurring patterns that could be built within Coda and figured out how no-code makers could build AI Agents using just Tables, Formulas, and Buttons in Coda.

So now I am repeating this pattern with the broader Claude Code and Claude Cowork experiences.

The CodaClaw pack is for use from within a Coda Document; it uses the SH Go Agentic SDK to ‘execute’ the SKILLS and SCRIPTS stored in your Coda doc as a hierarchy of pages (just as you would use a folder hierarchy for Claude).

  • It uses the Coda MCP Server to read, update, and create Coda Docs in your Coda Workspace.
  • It can take real-world actions via any of the many Coda Packs and Sh Go Agents that you have added to those documents.
  • It is using your Coda Workspace as its long-term memory, RAG Knowledge Base, and learning system.

The GrammarlyClaw pack is almost identical but is designed to use the Grammarly Docs workspace as the primary user interface (but still stores everything in your Coda Workspace).

I am working on the SuperhumanClaw version as well, which will use the Superhuman Go user interface (so it is available everywhere) but still uses your Coda Workspace Documents as above.

Because they use the Coda MCP Server and the Superhuman Go Agentic SDK facilities, they are (for now) only available to those who are part of the Beta Test Programs of those tools. (Please DM me if you are part of those programs and interested in participating in the Alpha Testing of these packs, and influence how they evolve.)

Meantime, my message to my clients is this:

when Coda launches both the MCP Server and the new Pack SDK into full production,

Coda will become the most powerful and most accessible alternative to all these 'Lobster-like" systems - safe - powerful - better integrated - and no-code maker friendly!

and you can use all your existing Coda skills (Tables, Formulas, Buttons, Automations, Packs, etc) to create your own powerful AI Agent swarms.

What a time to be a Codashian!

respect
Max

3 Likes
I made a 'thing' that might be useful?
2

I think it’s time to pitch Ryan Seacrest on a new reality series

1 Like
3

UGHHH - Dad joke alert!!!

4

I cannot wait for the “no-code maker friendly” portion.

5

Several people have asked me about the implementation of this “CodaClaw” thingy.
If OpenClaw is so dangerous (if you are not a CLI using dev), then how is CodaClaw any safer?

The issue is that OpenClaw runs on your PC and has read/write/delete access to EVERYTHING on your system, and can use your credentials to log in to and access EVERYTHING on the internet that you can personally access.

And sometimes it hallucinates or gets too proactive and can do bad things before you can stop it.

But CodaClaw is sandboxed inside your Coda workspace. You specify on the SKILLS pages which pages, tables, other documents, and coda packs are allowed to be read/write/deleted. You also provide all the REFERENCE materials and ASSETS in your document as pages also.

Essentially, all these new ‘Lobster Bots’ (Claude Cowork, OpenClaw, etc) work the same:

There is an innermost loop for each STEP that does the following:

  1. reads your information hierarchy to build a PROMPT for this step
    (a) information about your preferences, the agent’s purpose, and the rules of engagement
    (b) list of TOOLS available for doing stuff and taking actions
    (c) the current SKILL (task description) being executed
    (d) a list of the steps the LLM wants to execute and their current status
    (e) the chat history so far so the LLM can track where it is in the process
    (f) some ‘memory’ objects that are persisted across sessions
  2. sends this complicated prompt to the LLM to be ‘executed’
  3. the LLM processes the request and returns its next response, which is either:
    (a) a TOOL to be executed (an executable deterministic code/formula that does stuff)
    (b) or a final result to be sent to the user
  4. if the response was a TOOL, then it is executed and the result is added to the chat history
  5. if the TOOL ran ok, then the loop is repeated until the result is a text block

So for the CodaClaw, you provide a list of BUTTONS, FORMULAS, and PACKS for the TOOLS
And a description of the PAGES, TABLES, COLUMNS etc that can be used to create the prompts

But, unlike OpenClaw, the CodaClaw engine cannot access your disk drive, or your bank account, or the internet as it likes - it can only use the BUTTONS and PACKS you provide.

So ordinary users can create Self-Directed AI Agents by writing structured instructions (SKILLS) in plain English (or any language) and have them executed reliably.

At least that is the goal. Still in the early stages of building and testing all this.

Max

3 Likes
6

I tend to avoid calling the things that direct agents skills because a given agent may have several skills. I think these are more like agent workflows which almost always use one or several skills. The nuance is subtle, but imagine a workflow that needs to find a record of type in a table named . A skill that does that is utilized in a workflow and it may be used by several workflows and by several agents. The workflow guides the agent; the agent is given skills; skills help the agent achieve its workflow-defined goals.

CLAWS all the way down! And now, there’s NemoClaw, the enterprise, secure, SOC2-compliant claw from NVIDIA. Who knew we’d be here just a few months ago?

1 Like