Codegraff
macOS, Windows & Linux
Work with agents on code, research, and plans, with files, browser, and terminal in one place.
Explore CodegraffMore intelligent AI harnesses
Better AI. Less waste.
We build more intelligent AI harnesses, so models can do useful work with less wasted time and cost.
Products
Put AI to work.
From the first idea to the finished task. Work with AI in your workspace and across the web.
Browse
macOS
An AI agent, built into a browser you’ll love using. Research a question, work with a page, or fill in a form with Codegraff. All in a beautifully minimal Mac browser you can make your own.
Explore Browsenanoterm
macOS · open source
A terminal of blocks with a coding agent beside you. Your own shell, with Codegraff running commands as blocks next to yours, and native visuals in the answers.
nanomon
macOS · open source
A small, smooth activity monitor. Every process, and your CPU, memory, GPU, disk and network over time, with Codegraff in the corner to suggest what to quit.
The approach
Make every step count.
An AI harness is the system around a model: its tools, context, instructions, and checks.
- 01
Find the right context
Find the information the task needs without flooding the model with every file.
- 02
Choose a useful route
Match the task to useful models, tools, and a sensible budget.
- 03
Check the work
Test the finished work, record what it cost, and use that outcome to guide the next run.
Built with
Research
The thinking behind the tools.
Explore how agents find context, choose a way forward, and check their results.
Research in progressBetter harnesses. Measured in real work.Published Graff evaluations, task outcomes, costs, and runtime findings, with the conditions behind each result.Explore the evidence
Our Graff FrontierHarness evaluation brings the results, costs, and test conditions together.
Methods notes
All research notes