Introduction
context-mode is an MCP plugin that keeps raw data out of the LLM context window — run code in a sandbox, search a knowledge base, and only the answer enters the conversation.
Large language models reason inside a fixed context window. Every log line, API response, and file you read into that window is reasoning capacity you can't get back. A single unrouted command can drop tens of kilobytes of raw text into the conversation — and once it's there, it costs you for the rest of the session.
context-mode is an open-source MCP plugin that solves the other half of the context problem: instead of pulling raw data in, it runs the work out — in a sandbox — and returns only the derived answer.
The model writes a small program that processes the data and prints a summary.
The raw bytes stay in the sandbox; only what you console.log() enters context.
This is the Think in Code principle, and it routinely saves 95–99.9% of a
payload's footprint.
How it works
context-mode wraps four mechanisms behind a handful of ctx_* tools:
- Sandboxed execution — run code in 12 languages; only stdout returns.
- An FTS5 knowledge base — index large outputs and recall them on demand with BM25 search, so nothing is read twice.
- Web without the flood — fetch and index pages; raw HTML never enters context.
- Session continuity — decisions, errors, and plans survive compaction, so a resumed session already knows what you were doing.
Start here
Quickstart
Install and route your first command in a few minutes.
The tools
The complete ctx_* reference, with parameters and examples.
Core concepts
Think in Code, context protection, and the tool-selection hierarchy.
Platforms
Install on Claude Code, Codex, Gemini, Copilot, and more.
Why it matters
A session that would fill its window in ~30 minutes can stay productive for hours. You keep the same answers and the same work — at a fraction of the token cost — and your agent never loses the plot to a wall of raw output.