context-mode: 98% less tool output for AI agents

TL;DR: context-mode is a TypeScript layer for AI coding agents that sandboxes tool output, persists session memory, and routes calls across 17 platforms. The project claims a 98% output reduction on its own description, which is worth treating as a vendor number until you measure it.

What it claims

The description is dense and specific: “Context window optimization for AI coding agents. Sandboxes tool output (98% reduction), persists session memory, and enforces routing across 17 platforms via MCP + hooks.”

Three mechanisms, one goal. Sandboxing tool output means an agent’s file reads and command results do not flood the transcript. Session memory means state survives across turns rather than living only in the context window. Routing means the same setup drives whichever agent platform you happen to be using that day.

Both numbers worth scrutinizing live in that description: the 98% reduction and the 17 platforms. Neither is independently verified here. They are the project’s own claims, stated by its authors, about their own tool. Treat them as a starting hypothesis and measure against your own transcripts, because the real reduction depends entirely on how verbose your tools already are.

The routing claim is more falsifiable than the percentage, and it is the one you can check in an afternoon.

Where it hooks in

GitHub reports these topics, and they describe the integration surface precisely:

• claude, claude-code, claude-code-hooks, claude-code-plugins, claude-code-skill
• antigravity

The Claude Code entries outnumber everything else in the list, which tells you where most of the attention sits. Hooks are the mechanism that makes the routing stick, since a hook can intercept a tool call before it reaches the model and decide what the model actually sees.

The home page is context-mode.com.

Field	Value
Stars	24,793
Forks	1,784
Language	TypeScript
License	Not recognized by GitHub
Created	2026-02-23
Last push	2026-10-02
Size on disk	32 MB

The license field deserves a flag. GitHub returned NOASSERTION, which means it found a license file but could not map it to a standard SPDX identifier. That is different from no license at all, and it often means a custom or non-standard grant. Read the file before you depend on it commercially.

The repository received a push on the day this was written.

Before you install it

This is a layer that sits between your agent and its tools. That is a powerful position and it is worth being deliberate about it:

• Anything that sees your tool output sees your source code, your file paths, and anything else your commands return
• Hooks run on your machine, so you are installing executable code into the critical path of your editing loop
• The 98% figure is self-reported and unreplicated here

Cloning to read before you wire it in:

git clone https://github.com/mksglu/context-mode

At 32 MB it is large enough that a real read is worth doing, especially given where it sits in your workflow. Start by confirming what it logs and what it retains.

When it fits

It fits if your sessions are genuinely context-bound. That is a real and common failure mode, and the symptom is obvious once you have seen it: the agent keeps re-reading files it already read, and your per-task cost climbs for no gain.

It does not fit if your tool output is already terse. Sandboxing something that was never verbose adds a layer of indirection for no gain.

Judge it on your own transcripts. The claim is plausible, the mechanism is reasonable, and the 98% is not a number you should take on faith.

Source and Accuracy Notes

• Project page: https://github.com/mksglu/context-mode
• Source repository: mksglu/context-mode (verified via GitHub API html_url)
• License: NOASSERTION, GitHub found a license file it could not match to an SPDX identifier
• Stars at time of writing: 24,793
• The 98% reduction and 17 platform figures are claims from the repository description and were not independently benchmarked. Counts, topics, and dates were read from the GitHub REST API.