Skip to content

tarek-clarke/RedClaw

RedClaw: Trustworthy, Local-First Browser Agent for Job Applications

RedClaw is a production-grade, privacy-first autonomous agent designed to handle the complexity and sensitivity of job applications. Engineered specifically for AMD GPUs (ROCm) and LM Studio, it puts the human in control with a robust "Action Plan" and "Audit Log" architecture.

Why RedClaw?

In a world of black-box AI agents, RedClaw prioritizes trust, transparency, and local execution.

  • Local-Only: Your profile, resume, and credentials never leave your machine. All reasoning happens via your local LM Studio instance.
  • Human-in-the-Loop (HITL): RedClaw acts as a co-pilot. It proposes action plans for your approval and escalates ambiguous fields to you in real-time.
  • Auditability: Every run generates a structured JSONL audit log, capturing every observation, decision, and approval.
  • AMD Focused: Optimized for Radeon hardware (e.g., 7900 XT) via local GGUF models (Gemma 4 & gpt-oss-20b).

Key Features

  • Action Plan Preview: Approves a human-readable roadmap before a single browser action is taken.
  • Dry-Run Mode: Simulate full application flows without clicking final submission buttons.
  • Safety Policy Layer: Configurable domain whitelists and mandatory pauses for logins/uploads.
  • Site Adapters: Modular support for major hiring platforms (Greenhouse, Lever, etc.) for high-precision form filling.
  • Job-Fit Preflight: Automatically scores how well your resume matches a job description before starting.
  • Persistent Sessions: Securely reuse browser logins for approved sites without re-authenticating.

Security & Trust Model

Feature Behavior
Data Storage All personal data (user_profile.json, resume.pdf) is strictly local and Git-ignored.
Connectivity 0% Cloud dependency. Connects only to your specified LM Studio host.
Submission Never clicks "Submit" without explicit human approval.
Audit Logs Structured logs record all agent actions, screenshots, and user approvals.
File Access Restricted to the application directory and specific whitelisted local paths.

Quickstart

1. Requirements

  • Python 3.10+
  • LM Studio running with Gemma 4 (Vision) and gpt-oss-20b (Reasoning).
  • (Windows users) Double-click run.bat
  • (Mac/Linux users) Run ./run.sh

2. Configure Your Profile

Rename user_profile.example.json to user_profile.json and add your career highlights and links.

3. Usage

python main.py --goal "Apply for the Senior Machine Learning Engineer role at [Company Name]" --url "[Link]"

Dry-Run Mode

Use the --dry-run flag to test an entire application flow. The agent will navigate, extract, and fill data but will always pause before final submission.

Architecture

RedClaw uses a modular Adapter-Strategy pattern:

  1. Preflight: Scores job-fit and prepares answers.
  2. Planning: Proposes a multi-step roadmap.
  3. Execution: Uses Playwright with site-specific adapters for high reliability.
  4. Audit: Logs every step to logs/run_ID.jsonl.

RedClaw is built by the community for private, ethical AI automation.

About

Local, AMD-supported autonomous browser agent. Privacy-first task automation via LM Studio and Gemma 4.

Topics

Resources

License

Contributing

Security policy

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors