open-source MCP plugin · context engineering

The other half of the context problem.

context-mode keeps raw data out of the LLM context window. Run code in a sandbox, search an FTS5 knowledge base, and only the answer enters the conversation — so your agent stays sharp for hours, not minutes.

Get startedExplore the toolsView on GitHub
98%
context saved
17
agent CLIs
12
languages
~30×
fewer tokens

§ 01 · the idea

Think in code, not in raw bytes.

Every byte a tool returns costs reasoning capacity for the rest of the session. Reading a 700 KB log spends 700 KB of your window on bytes you’ll never reference again.

context-mode flips this: the model writes a small program that processes the data in a sandbox and prints only the conclusion. The raw data stays out of context, indexed and searchable for when you actually need a detail.

Same answers. Same work. A fraction of the tokens.

§ 02 · the toolkit

A small set of sharp tools, each built to keep context clean.

ctx_execute

Run code in a 12-language sandbox. Only what you print enters the conversation.

56 KB → 299 B

ctx_execute_file

Read a file into the sandbox and derive the answer — the bytes never reach context.

45 KB → 155 B

ctx_batch_execute

Run many commands in parallel, auto-index every output, and search them in one round trip.

986 KB → 62 KB

ctx_search

Query a unified FTS5 knowledge base with BM25 ranking and fuzzy correction.

on-demand recall

ctx_fetch_and_index

Fetch the web, convert to markdown, and index it — raw HTML stays out of context.

curl, without the flood

ctx_stats

See exactly how much context window you saved, with a per-tool breakdown.

measure everything

§ 03 · how it fits

Three steps, then it disappears into your workflow.

01

Install once

Add context-mode to your agent CLI with a single command. No accounts, no cloud.

02

Work as usual

Large tool outputs — logs, API responses, snapshots — are intercepted and routed to a sandbox.

03

Only answers enter context

The model sees the derived result, not the raw bytes. Your window stays clean for hours.

§ 04 · the numbers

Real payloads, measured.

See the full benchmarks

Playwright snapshot

56.2 KB → 299 B

99.5%

GitHub issues (20)

58.9 KB → 1.1 KB

98.1%

Access log (500)

45.1 KB → 155 B

99.7%

Git log (153)

11.6 KB → 107 B

99.1%

§ 05 · everywhere you work

One plugin, every agent CLI.

The same engine plugs into the host you already use, adapting to each one’s hook model.

Platform support matrix
  • Claude Code
  • Codex CLI
  • Copilot CLI
  • Cursor
  • Gemini CLI
  • VS Code Copilot
  • JetBrains Copilot
  • Qwen Code
  • Kimi Code
  • OpenCode
  • OpenClaw
  • Pi
  • Antigravity
  • Antigravity CLI
  • Kiro
  • Zed
  • OMP

Stop paying for context you never read.

100% local. Zero telemetry. Install in a minute and watch your context window stop filling up.

Read the quickstartStar on GitHub