Context-mode claims a 98% cut in AI tool output

Context-mode cut one agent's tool output from 315KB to 5.4KB, a 98% reduction, per its GitHub page.

MCP tool calls routinely burn context window on raw JSON and file dumps before an agent does any real work. The project already has 19,893 GitHub stars and 1,430 forks, a sign plenty of builders feel that same pain.

The tool, released under the ELv2 license, sandboxes MCP output so raw data never enters the context window. It also logs every file edit, git operation, task, and error in a SQLite database. After a session compacts, it pulls the relevant history back with FTS5 and BM25 search.

Separately, it routes across 17 agent platforms and trims filler the model tends to add on its own.

The clearest example: a task that took 47 separate Read() calls and 700KB now runs as one ctx_execute() script returning 3.6KB.

Think in code

Context-mode calls this pattern Think in Code, and it doesn't require installing the tool to use. Instead of reading files one by one to find an answer, the model writes a script that computes the answer directly. That's the part worth copying into any agent workflow, MCP server or not.