[KongchangAI]
Product

Context Mode

Context Mode is a plugin tool based on the MCP protocol, designed for AI coding assistants such as Claude Code, used to manage and compress context information. Its core mechanism persists data returned by tool calls to a local SQLite database while retaining only distilled summaries in the active context, reducing context bloat in long-conversation scenarios and improving the sustained working capability of large language models on complex tasks.

Timeline (last 90 days)

Oct 14

Context Mode is an MCP Server that connects to tools such as Claude Code, Codex, Cursor, and Gemini CLI via Hooks, and the total number of compatible Agents reaches 18.

Unverified50%
Oct 14

Context Mode's core design is that commands are first executed in a sandbox, only the final result is returned to the Agent, and the remaining intermediate data is written to a local search database for the Agent to retrieve on demand.

Unverified50%
Oct 14

Context Mode's official example shows 315 KB of context compressed to 5.4 KB, with session duration extended from about 30 minutes to a maximum of 3 hours.

Unverified50%
Oct 14

Hooks are interception points reserved in an Agent's execution flow (such as before a tool call, after a tool returns, and at session start), and Context Mode registers Hooks to intercept and rewrite tool output when it enters the context.

Unverified50%
Oct 14

In less than 8 months, Context Mode gained 28,000+ stars, nearly 2,000 forks, 111 contributors, and more than 600,000 NPM downloads.

Unverified50%
Oct 14

On Hacker News, Context Mode received 570 points, climbed to 4th place on GitHub Trending this week, and gained nearly 1,000 stars in a single day.

Unverified50%
Oct 14

Context Mode's claimed 98% context reduction has no independent third-party verification; its self-tested Benchmark averages 96%, with a worst case of only 13%, and users generally perceive it as between 30% and 40%.

Unverified50%
Oct 14

Under some configurations, the Context Mode tool itself consumes about 15,000 Tokens per request, which may offset the savings on small tasks; some users also report that the model guesses more because it can only see fragments.

Unverified50%
Oct 14

Context Mode's fundamental trade-off is exchanging 'the model cannot see the whole picture' for context efficiency, which may not be worthwhile for tasks requiring global reasoning (refactoring, cross-file debugging, architecture analysis), while being almost imperceptible in local tasks (such as checking log errors).

Unverified50%
Oct 14

Since March, Context Mode has switched from a traditional open-source license to the Elastic License.

Unverified50%

1 more timeline events

All Facts (20)

Unverified

Context Mode将MCP工具调用的原始返回数据存储在本地SQLite数据库中,上下文只保留精炼后的摘要

90%
Unverified

Context Mode的98%上下文节省主要针对内置工具,第三方MCP工具的输出无法有效拦截

90%
Unverified

Context Mode的快照机制在压缩发生前将任务状态(包括规划树、文件修改记录、决策上下文)序列化保存,压缩后再注入新的上下文

85%
Unverified

Context Mode与Claude Code自带的Compact功能是互补关系:Compact做语义摘要压缩工具定义,Context Mode做索引压缩工具输出

85%
Unverified

Context Mode使用Reciprocal Rank Fusion算法合并两种搜索策略的排序结果

85%
Unverified

Context Mode使用BM25算法检索相关内容,生成不超过2KB的Session Guide

85%
Unverified

Context Mode支持12个平台,在所有平台上强制执行'用代码思考'的规则

80%
Unverified

Context Mode对Playwright截图的上下文压缩率达99%,访问日志压缩率达100%,7.5MB的JSON API响应压缩到0.9KB

80%
Unverified

Context Mode uses an SQLite database and the FTS5 full-text search engine to track 26 types of events, including file reads and writes, Git operations, task status, error messages, and user decisions.

60%
Unverified

After installing commonly used MCPs, 70% of the context has already been consumed before any work begins.

60%
Unverified

Context Mode is suited to tasks whose context naturally balloons, such as handling very large logs, browser snapshots, and large codebase analysis, and is less suited to small-scale, short-session everyday coding tasks.

50%
Unverified

Context Mode is an MCP Server that connects to tools such as Claude Code, Codex, Cursor, and Gemini CLI via Hooks, and the total number of compatible Agents reaches 18.

50%
Unverified

Context Mode's core design is that commands are first executed in a sandbox, only the final result is returned to the Agent, and the remaining intermediate data is written to a local search database for the Agent to retrieve on demand.

50%
Unverified

Context Mode's official example shows 315 KB of context compressed to 5.4 KB, with session duration extended from about 30 minutes to a maximum of 3 hours.

50%
Unverified

Hooks are interception points reserved in an Agent's execution flow (such as before a tool call, after a tool returns, and at session start), and Context Mode registers Hooks to intercept and rewrite tool output when it enters the context.

50%
Unverified

In less than 8 months, Context Mode gained 28,000+ stars, nearly 2,000 forks, 111 contributors, and more than 600,000 NPM downloads.

50%
Unverified

Context Mode's claimed 98% context reduction has no independent third-party verification; its self-tested Benchmark averages 96%, with a worst case of only 13%, and users generally perceive it as between 30% and 40%.

50%
Unverified

Under some configurations, the Context Mode tool itself consumes about 15,000 Tokens per request, which may offset the savings on small tasks; some users also report that the model guesses more because it can only see fragments.

50%
Unverified

Context Mode's fundamental trade-off is exchanging 'the model cannot see the whole picture' for context efficiency, which may not be worthwhile for tasks requiring global reasoning (refactoring, cross-file debugging, architecture analysis), while being almost imperceptible in local tasks (such as checking log errors).

50%
Unverified

Since March, Context Mode has switched from a traditional open-source license to the Elastic License.

50%

Source Articles