Carve
An open-source browser agent powered by large language models, supporting local models, OpenRouter, OpenAI, and Anthropic as backends. It enables automated web page interaction through natural language instructions and includes a built-in human confirmation safety mechanism for verifying actions before execution.
Timeline (last 90 days)
Carve is an open-source computer use agent that can operate any website using only text instructions, without needing to write specialized scripts or rules for each page.
The Carve project has been open-sourced on GitHub, with the repository located at github.com/getcarve/carve.
Carve supports local models, OpenRouter, OpenAI, and Anthropic models simultaneously.
Carve is trained to pause and confirm with the user before performing consequential actions such as submitting forms or sending messages.
The demo scenario showcased by the Carve developer involves having it automatically fill out a supply chain form, with no manual field-by-field data entry required throughout the process.
Carve's ultrafast mode has not yet been officially released, and the author has stated that official support will be added in a future version.
LLM-driven browser agents dynamically determine operation targets by analyzing visual screenshots or DOM structures of pages; compared to RPA they are more robust to page changes, but introduce new challenges such as model hallucinations, operational latency, and inference costs.
Speed and reliability often require trade-offs, and how confirmation mechanisms coexist with speed optimization in high-speed execution modes is a key measure of the practicality of browser agents.
All Facts (8)
Carve is an open-source computer use agent that can operate any website using only text instructions, without needing to write specialized scripts or rules for each page.
50%UnverifiedThe Carve project has been open-sourced on GitHub, with the repository located at github.com/getcarve/carve.
50%UnverifiedCarve supports local models, OpenRouter, OpenAI, and Anthropic models simultaneously.
50%UnverifiedCarve is trained to pause and confirm with the user before performing consequential actions such as submitting forms or sending messages.
50%UnverifiedLLM-driven browser agents dynamically determine operation targets by analyzing visual screenshots or DOM structures of pages; compared to RPA they are more robust to page changes, but introduce new challenges such as model hallucinations, operational latency, and inference costs.
50%UnverifiedSpeed and reliability often require trade-offs, and how confirmation mechanisms coexist with speed optimization in high-speed execution modes is a key measure of the practicality of browser agents.
50%UnverifiedThe demo scenario showcased by the Carve developer involves having it automatically fill out a supply chain form, with no manual field-by-field data entry required throughout the process.
50%UnverifiedCarve's ultrafast mode has not yet been officially released, and the author has stated that official support will be added in a future version.
50%