# LocalAgentBrowser > An autonomous web research agent driven by a local language model (LM Studio) and Playwright. Give it an > objective; it searches the web, reads pages, records findings with their sources and writes a cited report, on > your own computer. Local by default and built for small models on a CPU (four tested), with no GPU or API key > needed; optional cloud models through OpenRouter with your own key. Key facts: - Install: `pip install localagentbrowser`, then `python -m playwright install chromium`; run `localagentbrowser`. Needs Python 3.10 or later and LM Studio running a model (developed with Qwen2.5-7B-Instruct, 8,192-token context). - Good goals: questions you would research by reading pages. Facts, lists with conditions, comparisons with a recommendation, how-tos, explanations, views from several sources, lookups in official documentation. - Not yet: anything behind a form, a login or a button. It reads pages only; it does not click, fill in forms or sign in, so flight or hotel prices, shopping and booking sites won't work. - Speed: on a CPU a run takes about 10 to 60 minutes, and a small model sometimes returns thin findings. - Tested on four models with identical code (1.64, seven tasks, 90 minutes each, CPU only): Qwen2.5-7B finished 6 of 7, Llama-3.1-8B 7 of 7, Qwen2.5-14B 5 of 7, Qwen2.5-3B 3 of 7; no finished answer filled a gap from the model's own knowledge. - Windows: an installer (LocalAgentBrowser--setup.exe on the GitHub releases) with its own Python, a choice of the four tested models during setup, and Start-menu shortcuts; the report opens by itself when a run ends. - Cloud models (optional, since 1.65): through OpenRouter with the user's own key; much faster, but prompts and pages read leave the computer. On 1.66 Phi-4, GPT-4o-mini, Llama-3.1-8B and Gemma-3-4B each finished 7 of 7 tasks in 4 to 24 minutes for $0.04 to $0.14 a suite. On SimpleQA, Phi-4 alone answered 1 of 50; with LocalAgentBrowser 1.66, 45 of the same 50 and 170 of the standard 200 (85%). The Windows installer sets them up; otherwise set LMSTUDIO_BASE_URL to https://openrouter.ai/api/v1, LMSTUDIO_MODEL to an OpenRouter model (microsoft/phi-4) and LMSTUDIO_API_KEY to the key (the README's FAQ has the full settings). - Honesty by design: rules in code check the model's claims. Findings must be on the page shown. Before finishing, each part of the objective is checked against the findings. What was not found is reported as not found, and things the pages didn't prove are marked ("not shown in the pages read to be open-source"). - Privacy: searches, written from your objective, go to DuckDuckGo, and pages are read from the web like any browser; by default the objective itself, the model and the report stay on your computer (with cloud models turned on, the objective, the prompts and the pages read go to OpenRouter and the model's provider). - L.A.B. (LocalAgentBrowser) follows LAB COAT, Citations Over Assumed Truth: every line of an answer should come from a page it read, with the source, rather than from what the model assumes. - License: MIT. ## Docs - [README](https://github.com/SnowFXXX/LocalAgentBrowser/blob/main/README.md): what it does, the four models tested, install, first run, settings, FAQ - [How it works](https://github.com/SnowFXXX/LocalAgentBrowser/blob/main/docs/how-it-works.md): the loop, code layout, the checks that keep a small model on track, performance, limitations - [Results by kind of task](https://github.com/SnowFXXX/LocalAgentBrowser/blob/main/docs/task-results.md): single runs and the regression suite by version - [Example reports](https://github.com/SnowFXXX/LocalAgentBrowser/blob/main/docs/examples/README.md): three unedited reports from the 1.66 test run (Phi-4), with what each got right and wrong - [Changelog](https://github.com/SnowFXXX/LocalAgentBrowser/blob/main/CHANGELOG.md): every version and why it changed - [Regression history](https://github.com/SnowFXXX/LocalAgentBrowser/blob/main/docs/regression-history.md): each version measured on seven kinds of task - [SimpleQA results](https://github.com/SnowFXXX/LocalAgentBrowser/blob/main/benchmarks/simpleqa/README.md): the SimpleQA runs (50 and 200 questions), with every answer and grade - [SimpleQA dataset](https://huggingface.co/datasets/SnowFX/localagentbrowser-simpleqa): the same answers and grades on Hugging Face - [Technical paper](https://snowfxxx.github.io/LocalAgentBrowser/paper.html): Verification over Trust; the design, a four-model evaluation and where the measures failed ## Optional - [Project site](https://snowfxxx.github.io/LocalAgentBrowser/): findings, the four models, install and the challenge - [PyPI](https://pypi.org/project/localagentbrowser/): the package - [Article](https://snowfxxx.github.io/LocalAgentBrowser/article.html): We Can Do Better: A Small AI That Can't Cover Up - [Article](https://dev.to/snowfxxx/teaching-a-7b-model-on-a-cpu-to-finish-what-it-starts-1amc): Teaching a 7B model on a CPU to finish what it starts