okidoki: an ai teammate in your cursor!

Inspiration: The Problem

My dad manufactures and exports machines. My uncle runs a customs brokerage with 40 employees. Growing up, I saw businesses spend countless hours filling out the same forms, moving data, and repeating identical workflows.

The regulations are public, the forms are standardized, yet the work is still manual. And this problem exists everywhere, from customs and logistics to finance, insurance, and HR.

The Solution

okidoki is an AI teammate that operates software and gets work done for individuals and businesses.

We built two experiences:

Spaces: Think Slack, but with an AI teammate that actually works. Teams share a workspace, company vault, and an agent running on a private Mac VM. Mention @okidoki, and it can operate software, execute terminal commands, use programmatic tools, and complete workflows efficiently.

Shake to Summon: Shake your cursor to summon okidoki. It can teach you software, guide you through tasks, or simply do the work for you.

The magic? Replay. Teach okidoki a workflow once, and it can repeat it using deterministic automation, only calling an LLM when necessary. This makes repetitive tasks dramatically faster and cheaper.

Why We're Different

We're not another ChatGPT wrapper. Our hybrid architecture combines AI reasoning, programmatic tooling, and deterministic automation instead of relying on an LLM for every click.

3 seconds to complete an entire U.S. customs entry form using replay.
Faster than browser-use (80K GitHub stars) in our internal benchmarks.
Minimal LLM dependency: Replay runs programmatically whenever possible, using AI only when needed.
Built for different workflows, from customs filings to accounting, HR, insurance, and enterprise operations.

How We Built It

We used Claude for reasoning, Apify for connectors, Convex for our backend and real-time collaboration, Moss for indexing, and Cartesia for voice, alongside our custom computer-use engine and replay framework.

Our biggest challenges were speed, reliability across software, and security. We tackled these through action batching, intelligent waits, approval systems, and hybrid execution.

Our biggest learning? AI doesn't need to reason through every click. Combining traditional software engineering with AI makes agents significantly faster, cheaper, and more reliable.