A note from the founder
More help with company work.
The final say stays with you.
Shiba is an AI employee workspace for the work behind a company: engineering, design, marketing, sales, operations, finance, and research. Describe the task, approve the proposed work, and review what comes back.
01 / Where it started
I wanted more choice in how agents worked.
I started with coding agents. I used Devin alongside local tools and wanted to choose the model for each task, run the software in my own account, and inspect the result.
I wanted Devin-level background execution — with my own tools, my own routing, and a clear point where I could say yes or no.
That became Shiba. The aim is broader now: help with the work around a company, from a code change to the brief, research, and operating notes that go with it.
02 / The work
Start with a useful piece of work.
An employee is a working brief: what to do, what context to use, and what to hand back. These are examples of responsibilities you can give one.
- Engineering
- Fix a bug, add a test, and explain the change.
- Design
- Draft a user flow, interface copy, or a design specification.
- Marketing
- Turn product notes into a launch brief and editable campaign copy.
- Sales & revenue
- Prepare an account brief, outreach draft, or summary of supplied pipeline notes.
- Operations
- Write a release checklist, handoff, or recovery runbook.
- Finance
- Organize supplied expense information or prepare a draft budget for human review.
The built-in starting templates are Engineering, Marketing, Research, and Operations. Use custom instructions for other roles. A role does not connect your accounts or give permission to publish, send messages, update a CRM, or move money.
For now, every execution needs a GitHub repository, including document work. It provides context and a place for the files. You review the actual output before using it.
03 / Your decision
Keep the important decisions visible.
- Choose the scope. Give a concrete task and say what a useful result looks like.
- Approve the request. Read the repository, task, and publishing choices before standard repository execution starts.
- Review the result. Inspect the files, changes, and reported checks. A finished task still needs your judgment.
The conversation stays in a thread, so you can discuss the result without starting the context again. Another execution still follows its approval flow.
04 / Choosing a tool
Factual comparison: Devin Cloud vs. Shiba
Devin helped shape the original question, but Shiba is not a claim of feature parity or a benchmark result. Choose the tool that fits the work and the responsibility your team wants to take on.
Shiba's source is available under AGPL-3.0-only, and you can run it in your own Cloudflare account. The software is free to self-host; hosting and model usage cost extra. You also take responsibility for setup, access, and operation.
Try a small task you can check, then judge the output, the review effort, and the actual cost. The evaluation guide gives you a starting checklist.
Try one task before changing your workflow.
Shiba is still under development. Start with a low-risk task and review deployment readiness before using it for live work.
