From the team behind aidrop.it — one workspace to build, host, and keep changing your code.
If you keep hitting Claude or ChatGPT usage limits, the cause is usually not how much you work — it's how much your agent re-reads. Feeding whole files and full project context into every session burns tokens on text the task never needed. The fix is retrieval: store knowledge outside the context window and pull in only the relevant pieces. Same work, a fraction of the tokens — and usually better answers.
Where your limits actually go
Token spend is dominated by input, not output: every file the agent opens, every doc you paste, every CLAUDE.md line loads into context — and gets re-sent with every subsequent message in the session. A 50k-token context re-sent over twenty turns is a million tokens before the agent has produced anything. Long sessions also trigger compaction, which itself spends tokens summarizing.
Re-reading vs retrieving
Re-reading means the agent loads broad context "just in case": whole files, whole histories, whole wikis. Retrieving means knowledge lives in an indexed store and the agent searches it, pulling two or three relevant chunks per task. The difference per question is routinely 10–50×: a handful of 200-word chunks instead of tens of thousands of tokens of maybe-relevant text.
Accuracy improves too. Models attend better to short, relevant context than to long, diluted context, because more text triggers context rot — the "needle in a haystack" problem is real. Retrieval doesn't just cost less; it points the model at the signal.
What this looks like in practice
Keep per-repo files like CLAUDE.md short — stable conventions only, one or two screens. Move the long tail of knowledge — decisions, client context, project history, preferences — into a retrieval-backed memory. With a memory connected over MCP, agents query it and pay tokens only for the chunks that match the task. The memory can grow for years without growing your per-session cost.
Habits that save limits
Start fresh sessions instead of stretching one conversation across unrelated tasks — every message re-sends the whole history. Don't paste documents the agent can search instead. And when a session produces a durable fact, save it to memory: next time it's a 200-token retrieval, not a re-explanation or a re-read. Deciding what enters the window is a skill of its own — context engineering.
FAQ
Why does Claude Code use so many tokens? Mostly input: files read into context are re-sent with every turn of the session. Long sessions with large files compound quickly — fresh sessions plus retrieval-based context cut this dramatically.
Does a bigger context window fix the problem? No — it raises the ceiling and the bill. You pay for every token in the window on every call, whether or not the task needed it. Selection saves money at any window size.
Is retrieval less accurate than giving the model everything? Usually the opposite. Models reason better over short, relevant context than long, diluted context. Good retrieval finds the needle; a giant window just grows the haystack.
One workspace to build, host, and keep changing your code
The repository, what the project knows, the rules a change has to follow, and the path to an address — kept together, and reachable by the coding agent your team already uses.