got claude code access at work and immediately went down the rabbit hole of making it useful inside a heavily restricted environment. no curl, no outbound http, sandboxed file system, no easy apis to build custom integrations against. the built-in browser tool is surprisingly potent though. if you're already authenticated in your system browser, you can point claude at internal dashboards, wikis, or project trackers and it reads the rendered dom as text. no api keys needed. someone internally built a tool that does exactly this, scraping authenticated pages and dumping structured text for the model to reason over. the pattern is basically browse once, save to a local cache directory, then run analysis prompts against that static text.
i'm also comparing it with other tools. gemini with workspace extensions can search drive and email natively, plus it has live web search. claude code can't do either of those. so the stack i'm settling on is gemini for anything that needs workspace access or external research, and claude code for code generation and cross-referencing local files. for budget controls, i'm thinking about model routing defaults, prompt caching for reused context files, and project tagging so we can track burn by team. the dream is an automated morning digest that pulls calendar, email threads needing replies, meeting action items, and any metric anomalies from dashboards. not all the way there yet but the browser tool gets you surprisingly close without installing anything unapproved.