Dassi

Hey Indie Hackers! I'm Waylon, co-founder of Dassi.

Like many of you, I run repetitive browser workflows every week: reconciling payout reports from Stripe and marketplaces against Google Sheets, screening job applicant profiles, and aggregating analytics dashboards.

When browser agents started launching, they mostly took one shape: a remote headless browser booted in a datacenter, costing $200/month. But as soon as you try using them for real business work, they hit a wall:

1. The Login Wall: You have to enter your passwords and 2FA codes into a remote VM.
2. The Drift Problem: Taking screenshots and guessing (x, y) coordinates piles up compounding errors on multi-step workflows.

Why We Built Dassi Inside Your Own Chrome
Instead of booting a second browser in the cloud, Dassi runs directly inside the Chrome browser you already use. It uses your existing authenticated sessions and password managers—zero credentials shared, zero duplicate logins.

Under the hood, Dassi uses the Model Context Protocol (MCP) and structured accessibility trees rather than raw screenshots. This lets it inspect exact DOM states and execute multiple actions in a single step.

The Benchmark Results
We recently put Dassi through the Carnegie Mellon Odysseys benchmark (200 real-world long-horizon web tasks rebuilt from actual user histories):

• 91.5% overall completion rate (183 / 200 tasks fully completed), the highest published score.
• 92.7% completion on the 109 'hard' tasks (multi-site errands with 8-12 strict checkpoints), compared to 11.0% for vision-only computer use models (Claude Opus 4.6).
• 96.3% on Ohio State's Online-Mind2Web (289 / 300 tasks).
• Runs on a Flash-class model (Gemini 3.8 Flash), costing ~$0.66 median per task rather than burning massive frontier API bills.

All task logs and eval scripts are open source:
• Benchmark repo: https://github.com/omnify-labs/dassi-evals
• Deep-dive writeup: https://dassi.ai/blog/dassi-odysseys-benchmark-highest-score/
• Website: https://dassi.ai

Would love to hear from other indie hackers: What recurring browser tasks eat up most of your week, and what would you want an AI browser agent to handle first?