Low Context Fixer

What 'running low on context' actually means

Claude Code has a finite context window - the total amount of text it can hold in mind at once, made up of your CLAUDE.md, the files it has read, the whole conversation so far, and the tool definitions from any MCP servers. When that fills up, you get the 'running low on context' warning. Past a certain point, answers get worse (older context is squeezed out, so it forgets things and repeats itself) and every turn gets slower and more expensive because the whole history reloads each time. The tool above tells you whether to compact or restart, and estimates what is filling your window.

The single most useful decision here is compact vs clear. Compacting summarizes the conversation so far and keeps you in the same thread - the right move when you are mid-task and want to keep going. Clearing wipes the slate for a fresh full window - the right move when the thread has drifted, is thrashing, or has gone down a wrong path that you do not want to carry forward.

Compact vs clear vs restart

• /compact - summarizes the current conversation and frees space while keeping the thread going. Use it when you are mid-task and just need more runway.
• /clear - resets the conversation to a clean, full window. Use it when the thread drifted, is repeating itself, or you are starting a genuinely new task.
• Restart the session - like /clear but from a fresh process. Use it when the session feels heavy even after compacting, or you want a clean slate for something unrelated.
• Trim inputs - shrink CLAUDE.md, load fewer files, disconnect unused MCP servers. This reduces the baseline cost of every turn, not just this one.

What actually fills your window

1. Conversation history - the biggest growing cost. Every turn adds to it, and the whole thing reloads each turn. Long sessions are the number-one cause of context pressure.
2. Files read into context - a single large file read in full can eat a large chunk of the window. Half of context blowups are 'I asked it to read this big file and it kept the whole thing'.
3. CLAUDE.md - loaded into every single turn, so every line is a permanent cost. A bloated CLAUDE.md taxes you constantly, not just once.
4. MCP servers - each connected server loads its tool definitions into context. A pile of MCP servers you are not using is pure overhead. Disconnect the ones this task does not need.

Preventing the warning in the first place

The best fix is not needing the fix. Keep CLAUDE.md tight - under a couple hundred lines - because it loads every turn. Scope tasks so sessions stay short; when you finish one thing, clear before starting the next rather than letting one mega-session sprawl. Point Claude Code at file paths and let it read the specific parts it needs instead of pasting whole files. And disconnect MCP servers you are not actively using in this session. Do these four things and you will rarely see the low-context warning at all.

Reading the estimator

The estimator on the right converts approximate sizes into a rough token count using the standard back-of-envelope of about four characters per token. It is a planning aid, not an exact meter - the real counts depend on the exact content and the model. The useful signal is proportion: which bar is biggest. If loaded files dominate, stop reading whole files. If history dominates, compact or restart. If CLAUDE.md is surprisingly large, trim it, because you pay that cost on every turn forever.

Context discipline is a skill

The developers who get the most out of Claude Code treat the context window like a budget they actively manage - compacting at the right moment, clearing decisively when a thread goes sideways, and keeping their inputs lean by default. It is a learnable habit, and it is one of the highest-leverage things you can practice, because it improves answer quality and cuts cost at the same time. The Claude Code Club curriculum has a full module on context management and session hygiene for people doing serious work.