even a frontier model can write broken code when its guessing how your app works XERJ helps your agent find the facts it needs faster and more precisely. better context gives it fewer reasons to invent an API call or make the wrong assumption about your code free open source repo: here is what that looks like when Claude Code needs to add Stripe Checkout to an existing app the project already has routes, pricing and frontend code. Stripe has its own API requirements. the agent needs both to write a change that fits here is the workflow, step by step: > step 1 → give Claude Code the task: “Add Stripe Checkout to this app” > step 2 → XERJ finds the existing checkout route, server-side price configuration and frontend request > step 3 → the agent searches the Stripe reference pack for Checkout Session creation > step 4 → the relevant method and parameters come back alongside the project files, with source locations attached > step 5 → Claude Code uses those references to start writing the checkout route > step 6 → the proposed change follows the app's existing price mapping and response format knowing how to write a payment endpoint is only part of the task. the agent also needs to know where it belongs, which prices to use and what the frontend expects back. missing any of that can leave you with code that looks reasonable but breaks when you connect it to the app. XERJ makes those details searchable while the agent works. when another question comes up, it can look up the answer and check the source that can help a cheaper model handle unfamiliar code and help your expensive model make fewer avoidable mistakes. both get better material to reason from. XERJ runs locally and connects to coding agents over MCP full XERJ breakdown below ↓