Every month a new "agent platform" launches, and every one of them opens with the same form field: paste your API key. It looks harmless. It isn't. That one field quietly discards most of what makes your coding agent good.
What an API key doesn't carry
When you run claude in your terminal, you're not just getting a model. You're
getting the whole environment you've built around it:
- Hooks that gate dangerous commands and format on save.
- Skills that encode how your team deploys, migrates, reviews.
- Subagents and their orchestration.
- MCP servers wired into your infra — databases, browsers, issue trackers.
- Subscription billing — the flat monthly plan you already budgeted, with its rate limits and its predictable ceiling.
An API-key integration reimplements the model call and abandons the rest. Your hooks don't fire. Your skills don't exist. And you pay per token, on top of the subscription you're already paying. For heavy agentic work, metered API billing routinely costs multiples of a flat plan doing the same tokens.
Wrapping the CLI instead
hang4r takes the other path: it drives the
claude, codex, and cursor-agent binaries you already
have, as subprocesses, speaking their native streaming protocols. The app never sees an
API key, never proxies your traffic through anyone's cloud, and never writes to
~/.claude. If the CLI can do it, the wrapped session can do it — because it
is the CLI.
This has a few consequences we consider features:
- Zero new trust. You already trusted the CLI vendor with your code and your login. hang4r adds no third party to that list — it's a local app on your machine.
- Rate limits become a UI problem, not a surprise. Since you're on your own plan, hang4r shows your Claude 5-hour and weekly windows as live gauges, plus per-session token and cost counters. You always know how much runway is left before you dispatch one more agent.
- Vendor upgrades are free. When Anthropic ships a new model or a new CLI capability, it shows up in hang4r the day you update the CLI — there's no waiting for a platform to re-integrate.
The honest trade-off
The subscription-CLI model has real limits, and you should know them. You can't sidestep your plan's rate limits (the gauges just make them visible). Cursor is the exception in the landscape — its agents bill through Cursor's own plans, so hang4r drives its CLI but your Cursor subscription is still the meter. And a local-first app means your machine (or a host you SSH into) does the work — there's no vendor cloud absorbing long tasks overnight.
We think that trade is right for working engineers: predictable cost, full capability, nothing new to trust. Your subscription is the API. The window is the product.