Job hunting is a repetitive cycle of finding postings, tailoring resumes, and fighting with application forms. A new open-source tool called Handoff, built on Claude Code, aims to automate the tedious parts of this process while enforcing a strict boundary between AI drafting and human execution. By integrating Model Context Protocol (MCP) servers with a local SQLite database and a Playwright browser instance, Handoff handles discovery, CV tailoring, and form filling, but stops short of sending emails or submitting applications without explicit user approval.

The Architecture of Control

The core design principle of Handoff is that the AI assistant can draft and record, but cannot send, submit, or access secrets. This is not achieved through prompt engineering alone, but through architectural constraints. The assistant interacts with a fixed set of MCP tools that lack the capability to trigger irreversible actions like sending an email or pressing a submit button. Those final steps are isolated in a local dashboard listening on 127.0.0.1, accessible only to the human user. A database check further prevents any job state changes that bypass these rules, ensuring that the AI cannot hallucinate a submission or accidentally email the wrong contact.

Workflow and Data Integrity

Handoff operates through a series of slash commands within Claude Code, such as /discover for finding jobs and /tailor for generating CVs. The system relies on a 'facts' database where the user inputs their employment history, skills, and projects once. Every claim in the generated CV or cover letter must cite one of these pre-approved facts, preventing the LLM from inventing numbers or experiences. The tool also provides an ATS-style score for each CV, highlighting missing keywords. For application forms, the /apply command uses a shared Chromium browser to fill fields, leaving the tab open for the user to review and submit manually.

Local-First Privacy and Setup

The entire workflow runs locally on the user's machine, with no data sent to a central Handoff server. All job history, drafts, and facts are stored in a single SQLite file, while sensitive credentials like mail logins are kept in an encrypted vault. Setup requires Node.js 20 or newer, a signed-in Claude Code instance, and a Chromium-based browser. The repository, available at github.com/DeazStar/handoff, includes a demo mode with fictional data to allow users to test the workflow without exposing their personal information. While currently optimized for developer roles on macOS, the author notes that Linux support is functional and Windows is untested.

Key Takeaways

  • Handoff uses architectural constraints rather than prompts to prevent AI from taking irreversible actions like sending emails.
  • The tool enforces data integrity by requiring every CV claim to cite user-provided facts, reducing hallucination risks.
  • All data remains local, stored in SQLite and an encrypted vault, with no external server dependencies.
  • The workflow is currently pre-release and tailored for developer jobs, with future plans for swappable packs for other professions.

The Bottom Line

Handoff represents the mature phase of AI adoption: moving beyond novelty to practical, constrained automation. By architecturally blocking irreversible actions, it solves the primary trust barrier in agentic workflows, making it a compelling model for how AI should integrate into high-stakes human processes.