context-mode Claims 98% Less Agent Context Under Elastic License
Software / news
context-mode Claims 98% Less Agent Context Under Elastic License
The 28,300-star plugin for Claude Code and about 17 other coding agents ships under ELv2, which Elastic says is not open source, and its README gives no method for the 98 percent figure.
context-mode, an MCP server and hook system that keeps raw tool output out of an AI coding agent's context window, has 28.3k GitHub stars and 2.0k forks. Its README says it turns 315 KB of tool output into 5.4 KB, "98% reduction," and the licence badge reads ELv2.
ELv2 is the Elastic License 2.0, and Elastic says in its announcement post that it is not an open-source licence by the OSI definition. The README never calls the project open source; the repository is owned by a GitHub account named mksglu, and the page gives no author name.
What the licence allows and forbids
Elastic's licence page lists three limitations. A user may not provide the software to third parties as a hosted or managed service that exposes a large share of its features. A user may not circumvent license-key functionality. A user may not remove licensing or copyright notices.
For a developer installing the plugin on a laptop, none of those bites. For a company that wants to host it as a team service, the first one does, and the README's own ctx_insight command already points to a hosted dashboard at context-mode.com/insight that it calls "org analytics for AI-assisted engineering teams."
The README does not say what data that dashboard receives.
The 98 percent figure comes with no method
The README says: "Sandbox tools keep raw data out of the context window. 315 KB becomes 5.4 KB. 98% reduction." It does not describe the workload, the tokenizer or how the number was measured. These are the project's own figures and the page shows no independent check.
It lists three examples with sizes: a Playwright snapshot at 56 KB, twenty GitHub issues at 59 KB, and a batch execution that falls from 986 KB to 62 KB. A separate ctx_execute example drops from 56 KB to 299 B.
- Batch execution, before986 KB
- Batch execution, after62 KB
- Overall example, before315 KB
- Overall example, after5.4 KB
Source: context-mode README, accessed 2026-10-14 (project-supplied, not independently verified)
How it works and where it can fail
The ctx_execute tool runs code in an isolated subprocess and passes only standard output back to the agent, across 12 language runtimes. Other content is chunked into a SQLite FTS5 index, searched with BM25 ranking and Reciprocal Rank Fusion. Hooks record file edits, git operations and errors, then rebuild session state after compaction.
The sandbox runs arbitrary code on the user's machine. The README describes process isolation but does not present it as a security boundary, and authenticated command-line tools such as gh, aws and kubectl inherit environment variables and config paths. The ctx_upgrade command pulls and rebuilds from GitHub.
The README also says hooks on Codex and Cursor can fail open or fire twice, and that Zed and Antigravity IDE have no hooks. It puts instruction-file-only compliance at about 60 percent, which is why it prefers hooks.
| Requirement | Value |
|---|---|
| Runtime | Node.js 22.5 or later, or Bun |
| Claude Code install | /plugin marketplace add mksglu/context-mode |
| Other clients | npm install -g context-mode, then MCP and hook config |
What the stars do and do not say
The plugin supports about 18 clients, including Claude Code, Gemini CLI, Cursor, Codex CLI, Zed and VS Code Copilot. Stars measure attention, a point this site made about REA, and the README gives no install count.
The case for the idea is separate from the case for this repository. A study covered here found autonomous coding agents raised commits by about 240 percent but releases by about 30 percent, which is why context discipline is a live concern for teams running them.
The repository's issue list showed one entry on Oct. 14: #45, opened March 3, 2026 by mksglu, a call for beta testers across 15 platforms and three operating systems, with 163 comments. A BENCHMARK.md file is listed in the repository,, and its contents were not read.
The check that would settle the 98 percent claim is to run the same agent task with and without the plugin and compare token counts, which the README does not do.
Sources
More in Software
- 01xtool Builds iOS Apps on Linux, With Gaps That MatterThe MIT-licensed Xcode replacement has about 9,400 GitHub stars, while its issue tracker shows WSL2 hangs, an empty widget binary and an open licence question.
- 02Gentoo Developer Asks Python Projects to Drop uv-buildMichał Górny says Astral's Rust build backend is 1.2 to 1.5 MiB against 45 KiB for flit-core, and that its release pace forces source rebuilds for Gentoo users.
- 03Apple's Safari-vs-Chrome Claims Rest on Prerelease BuildsThe footnotes on Apple's new Switch to Safari page show tests run in August 2026 on one laptop, against two different Chrome builds.
- 04Rust Borrow Checker Bug 11 Years Old Lets Safe Code Hide a Short-Lived Reference in a 'static ValueIssue #164250 shows a Drop impl smuggling a stack reference past the compiler on stable releases back to 1.85.0, and commenters trace variants to Rust 1.4.