LocalAgentBrowser
An open-source browser AI agent project that runs 7B-parameter local models on consumer-grade CPUs and autonomously completes search, reading, and writing tasks, using a single-file Python architecture.
Timeline (last 90 days)
LocalAgentBrowser is an open-source project that enables a 7B-parameter model to run a browser agent on a consumer-grade CPU and autonomously complete research and writing tasks.
The LocalAgentBrowser agent runs the Qwen2.5-7B-Instruct model on an Intel i5-9500 CPU and can complete the full workflow from searching and reading to writing an answer.
A typical run of the agent is: four turns of dialogue, completing an Africa's highest peak research task in about twelve minutes, with only search requests leaving the local machine throughout.
The entire agent is a single-file Python program, and in each turn the model receives the goal, information discovered so far, a log of recent actions, and the current page content, and outputs action instructions in JSON format.
The agent supports five actions: SEARCH, NAVIGATE, READ_MORE, DISCOVER and FINISH.
By maintaining a continuously growing conversation, appending only the changed parts, and adjusting the prompt order so that stable content comes first and changed content comes last, the agent reduced same-page turns from 150-220 seconds to 35-120 seconds.
One way to solve small models repeatedly making the same mistakes is to enforce checkable rules in code, including intercepting duplicate searches, refusing to open the page currently already on, blocking access to the same URL more than three times, and filtering blank and duplicate findings.
The duplicate URL check in early versions ignored the query string, causing every search after the first to be judged a duplicate, with 34 of 35 rounds blocked.
Three mechanisms help the small model learn to stop: a progress check after four consecutive rounds with no new findings, a goal check every eight rounds, and confirming whether content has been recorded before leaving a page.
Separating the writing step from the browsing conversation and using a fresh prompt containing only the findings can prevent the model from citing unverified page text or fabricating information.
All Facts (10)
Separating the writing step from the browsing conversation and using a fresh prompt containing only the findings can prevent the model from citing unverified page text or fabricating information.
50%UnverifiedThe LocalAgentBrowser agent runs the Qwen2.5-7B-Instruct model on an Intel i5-9500 CPU and can complete the full workflow from searching and reading to writing an answer.
50%UnverifiedLocalAgentBrowser is an open-source project that enables a 7B-parameter model to run a browser agent on a consumer-grade CPU and autonomously complete research and writing tasks.
50%UnverifiedThe entire agent is a single-file Python program, and in each turn the model receives the goal, information discovered so far, a log of recent actions, and the current page content, and outputs action instructions in JSON format.
50%UnverifiedThe agent supports five actions: SEARCH, NAVIGATE, READ_MORE, DISCOVER and FINISH.
50%UnverifiedBy maintaining a continuously growing conversation, appending only the changed parts, and adjusting the prompt order so that stable content comes first and changed content comes last, the agent reduced same-page turns from 150-220 seconds to 35-120 seconds.
50%UnverifiedOne way to solve small models repeatedly making the same mistakes is to enforce checkable rules in code, including intercepting duplicate searches, refusing to open the page currently already on, blocking access to the same URL more than three times, and filtering blank and duplicate findings.
50%UnverifiedThe duplicate URL check in early versions ignored the query string, causing every search after the first to be judged a duplicate, with 34 of 35 rounds blocked.
50%UnverifiedThree mechanisms help the small model learn to stop: a progress check after four consecutive rounds with no new findings, a goal check every eight rounds, and confirming whether content has been recorded before leaving a page.
50%UnverifiedA typical run of the agent is: four turns of dialogue, completing an Africa's highest peak research task in about twelve minutes, with only search requests leaving the local machine throughout.
50%