Theo is right. Bash is just the beginning. We built the "Execution Layer" he talks about in the video. We are calling it Worklayer. Here is a sneak-peek of the landing page. Details & product link in comments.

Worklayer is exposed to your agents via MCP. You can connect it to all the agents you already use: - Claude Code - Cursor - Codex - Cowork - OpenCode - Hermes - Pi w/ MCP Extension - Goose - your custom harness.

Every app you connect via Worklayer is available to your agents via code execution. Agents can write scripts spanning multiple apps to take action and get shit done.

The 3 huge benefits of using Worklayer: 1. Your agents work faster because they can write code to call multiple tools across multiple apps at once, instead of calling them individually.

2. Your agent consumes fewer tokens. We allow agents to control the output of the script, which lets them engineer context on the fly. Save costs on API plans; get more out of your Claude Code / Codex subscriptions.

3. The range of tasks your agents can complete increases significantly. Code gives agents the freedom to wrangle tools however they like. Filtering, sorting, and data transformations become easy.

Other benefits: - Avoid vendor lock-in - Centralized Vault for all your apps (like 1Password for agents) - Granular permissions: Allow, Ask, Deny We’re working on a lot more to help your agents self-evolve by incorporating skills and self-managing wiki. Much more to come!

Try Worklayer here: We just went live today. Please report any bugs you encounter. If you don't see your favorite app, DM me. I'll add support ASAP.

explain how Worklayer solves huge problems for users trying to use agents to get knowledge work done.

like @Cloudflare's code mode but for every app out of the box?

@Cloudflare Ding ding ding

Is this the "Code Mode" mentioned here?

Yes sir, but our own implementation. We don’t use cloudflare workers.

Nice! I will definitely try it!

Very cool! This architecture makes a ton of sense for provisioning access to external systems. I have 2 questions: - 77% token saving is *huge*. Is that all from just context mgmnt? - How do you evaluate the claim "agents complete 62% more tasks?" Congrats on the launch!

Token savings is very task dependant. You can obviously cherry pick data transformation heavy tasks to show upto 98% fewer tokens as done by Anthropic in this video: The 77% is measured by comparing Worklayer vs against official connectors from Anthropic. 62% more tasks calculation is done by using Claude Agent SDK on traditional tooling method vs Worklayer's programmatic execution method. We may potentially opensource our internal task library, but its extremely valuable eval IP and we want to think about it a bit before making it public.

it's open source?

Dang I didn’t realise vibecoding could be that quick

Is it like if composio and daytona had a sun? 🤔

Yes!

@sergical really like that scan animation

Typescript seems like such a least-common denominator solution to bash. All the bash criticisms are valid. But we’ve seen the mess the JS ecosystem is in with axios - and TS is an abstraction over that. I’m still thinking it through but Elixir feels like a good solution here, lightweight processes, easy to manage permissions for hierarchies of agents, a perfect agent toolkit in Jido…. A more structured execution layer than Bash is needed, but something powerful and flexible like the restructuring and pattern matching capabilities of the BEAM could work wonders here. Of course, @mikehostetler is probably already building it. 😂

yeahh core issue is bash has no failure semantics as exit code 1 means failed but not why' or should i retry' Once you need reliable agent orchestration, you want something with typedd errors and explicit retry contracts

excited to give it a spin! 🍁 quality home-grown software

legendary! congrats on the launch @trillhause_

Some connectors are broken we are actively fixing them asap. Expect all integrations to work by tomorrow.

This is the real gap. Agents need something between raw tool calls and full product workflows, otherwise every team keeps rebuilding the same execution, state, and recovery layer from scratch. Curious how opinionated Worklayer is about retries, traces, and permission boundaries.

Interesting framing. Letting agents compose scripts across multiple apps instead of calling tools one by one is a solid approach to the “bash is not enough” problem.

No local run is kinda a deal-breaker. A lot of people already have custom local MCP servers or network-protected APIs, looks like this doesn't fit in worklayer setup.

I feel like your site is a scam. Is it an scam?

No. Why do feel that?

When connecting my google integrations it sent me a message through google saying your a new app and flagged it

Calendar and forms are still under scope review form Google. We will remove them until approved. Thanks for flagging.

Of course my brotherr

I cant seem to connect it to opencode

Can you open DMs for me, I will help you debug it

worklayer looks clean execution layer idea finally getting its due.

Open source? If not, then you probably can’t compete