Developers spend a large share of their day navigating repository hosting interfaces: listing repositories, triaging issues, reviewing pull requests and collecting simple statistics. Large language models (LLMs) can turn plain-language requests into API calls, but giving a model write access to a live account raises obvious safety concerns. This case study presents RepoAgent, a web application in which a user connects a GitHub account once and then operates it through conversation. The system discovers every available GitHub operation at start-up, converts each operation's structured input schema into a JSON-Schema tool definition, and exposes one tool per operation to an LLM reached through an AI gateway. Each operation carries a risk level; any operation that is not read-only is placed behind explicit user approval, enforced in the server code rather than left to the model's discretion. Credentials are held per tenant by the Corsair integration layer, and users are identified by a lightweight cookie-backed session with no password. The interface streams the model's reasoning, shows every tool call as an inspectable card, and pauses at write operations with Approve and Deny controls. This paper describes the design, implementation and security properties of the prototype, defines an evaluation protocol, and analyses limitations, including the dependence of the guarantee on accurate risk metadata and the absence of real authentication.

Zamam Ansari, Manoj Singh, "RepoAgent: A Natural-Language, Approval-Gated Agent for Operating GitHub through Schema-Derived Tool Calling" International Journal of Scientific Research and Engineering Development, V9(5): Page(1365-1370) September - October 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.