How I Got an AI Agent to Open a Browser, Scrape Hugging Face Papers, and Auto‑Post to X

In this walkthrough I explore LocoAgent, an open‑source AI‑driven social‑media agent that operates through real Chrome browser sessions instead of platform APIs. The project (github.com/LocoreMind/locoagent) combines a local agent, lightweight models, browser automation, and a workflow engine to perform tasks such as liking, replying, following, and posting on X and LinkedIn.

Core Capabilities

Real browser automation: connects to an existing Chrome remote‑debugging port to reuse logged‑in sessions.

Platform skill system: defines commands like /x-com that encode platform‑specific actions for the agent.

Workflow engine: fixed scripts such as hf-daily-papers and hf-papers-to-x execute deterministically rather than relying on free‑form model output.

Operation log: records performed actions to avoid duplicate likes or posts.

The design mitigates two major risks of social‑media automation: uncontrolled model behavior and platform anti‑automation measures, by keeping the workflow stable and letting the agent focus on supervision and decision‑making.

Workflows

Key workflows include: hf-daily-papers: opens Hugging Face Daily Papers, extracts titles, links, abstracts, and thumbnails. hf-papers-to-x: downloads thumbnails and posts each paper to X. x-search-reply: searches X for keywords, reads posts, and generates replies with a model. linkedin-search-reply: same logic applied to LinkedIn.

Running bun run workflow list lists the four available workflows.

Customizing Workflows

Workflows are defined by JSON/TS files under the workflow directory; they can be edited or new ones added to automate any repeatable browser task.

Installation and Configuration

After cloning the repository, install dependencies with: bun install Create a .env file with OpenAI‑compatible settings, for example:

CLAUDE_CODE_USE_OPENAI=1
OPENAI_API_KEY=YOUR_KEY
OPENAI_BASE_URL=YOUR_COMPATIBLE_ENDPOINT
OPENAI_MODEL=YOUR_MODEL_NAME
SKIP_PERMISSIONS=1

Install the browser control tool and start Chrome remote debugging:

npm install -g agent-browser
agent-browser --version
bun run setup-chrome

The script copies a Chrome profile, launches the DevTools protocol on port 9222, and connects agent-browser to it, confirming with:

Chrome CDP ready
agent-browser is ready

Running a Workflow

Executing bun run workflow run --id hf-daily-papers opens Hugging Face. With the default filter (minimum 5 upvotes) no papers are selected, returning a partial result JSON indicating zero papers.

Lowering the filter to zero upvotes and re‑running captures 15 papers, fetching abstracts and saving data to workflows/.tmp/hf-2026-05-28/papers.json. The distinction is clarified:

Default workflow unchanged.

Temporary executor parameters relaxed for demonstration.

Resulting data stored locally.

The final X posting step also adds the paper link as a comment to align with X’s recommendation algorithm.

Conclusion

While LocoAgent can be used merely for automated tweeting, its real value lies in demonstrating how a local AI agent can orchestrate browser‑based automation, organize platform skills, structure workflows, log actions, and embed models into the pipeline. Researchers interested in local agents, browser automation, or building personal information‑flow pipelines should clone the repository and study its source.