MCP server · 17 platforms

Every tool call spends
the same budget
your thinking does.

A Playwright snapshot costs 56 KB. Twenty GitHub issues cost 59 KB. One access log, 45 KB. After 30 minutes, 40% of your context is gone — and when the agent compacts to free space, it forgets which files it was editing and what you last asked for. Context Mode sandboxes the raw data and indexes the session instead.

Read the source Install it

Context ledger — one session

figures as published

Playwright snapshot56 KB

Twenty GitHub issues59 KB

One access log45 KB

… and the rest of a working session315 KB total

Into the window315 KB

Sandboxed to5.4 KB

The lit sliver is what reaches the model — a 98% reduction, the project’s own figure. Every number on this panel is quoted, not measured here.

Four sides of one problem

the fourth is a refusal, not a feature

Context saving

Sandbox tools keep raw data out of the window. 315 KB becomes 5.4 KB.

Session continuity

Edits, git operations, tasks, errors and decisions are tracked in SQLite and retrieved by search — not dumped back into context. No --continue means previous session data is deleted immediately.

Think in code

The model should program the analysis, not compute it. One script replaces ten tool calls.

No prose-style enforcement

It never dictates how the model writes its final answer. The project cites evidence that aggressive brevity prompts degrade coding and reasoning benchmarks, and keeps its routing block focused on where data goes, not how the model talks.

Think in code — the project’s own example

// Before: 47 × Read() = 700 KB.  After: 1 × ctx_execute() = 3.6 KB.
ctx_execute("javascript", `
  const files = fs.readdirSync('src').filter(f => f.endsWith('.ts'));
  files.forEach(f => console.log(f + ': ' + fs.readFileSync('src/'+f,'utf8').split('\n').length + ' lines'));
`);