[KongchangAI]
Product

Billion Context

The most-downloaded plugin on the DSH platform. It controls context length by folding earlier conversation content into summaries, reducing the cost of repeated reading. It is suited to long-term use and scenarios spanning a wide range of topics.

Timeline (last 90 days)

Nov 9

The current version of Billion Context is 0.121; the project is evolving very fast and automatically updates the master branch.

Unverified50%
Nov 9

The code tab in the Claude desktop app bypasses the proxy, so it cannot take advantage of Billion Context's compression.

Unverified50%
Nov 9

When running with Billion Context's host option, there is no authentication mechanism, so it can only be used on trusted networks.

Unverified50%
Nov 9

Billion Context is a small proxy server that sits between an AI agent and a model API, managing the context window by reversibly compressing conversation history.

Unverified50%
Nov 9

Billion Context supports Anthropic- and OpenAI-style APIs.

Unverified50%
Nov 9

Billion Context's cache hit rate for healthy sessions can reach 95% to 97%, and the compression itself costs 2% or less.

Unverified50%
Nov 9

In the needle-in-a-haystack test, Billion Context's accuracy dropped from 69% to 27%.

Unverified50%
Nov 9

Billion Context can be integrated with clients such as Claude Code, OpenCode, and Codex, and supports models such as DeepSeek V4 Flash, Claude 3.1 Pro, GLM-5.3, Qwen 3.8 Max, and Minimax M3.1 Flash.

Unverified50%
Nov 7

In a 1.48-million-token test, full context failed all 199 tests, while Billion Context passed each time with a compression ratio of about 30x.

Unverified50%
Nov 7

Billion Context can be installed via npm (npm install -g billion-context), uses the billy command, and supports three integration methods: installing as a native plugin, launching the client directly through the proxy, and manually prepending the proxy address to the base URL.

Unverified50%

12 more timeline events

All Facts (20)

Unverified

The current version of Billion Context is 0.121; the project is evolving very fast and automatically updates the master branch.

50%
Unverified

The code tab in the Claude desktop app bypasses the proxy, so it cannot take advantage of Billion Context's compression.

50%
Unverified

When running with Billion Context's host option, there is no authentication mechanism, so it can only be used on trusted networks.

50%
Unverified

Billion Context is a small proxy server that sits between an AI agent and a model API, managing the context window by reversibly compressing conversation history.

50%
Unverified

Billion Context supports Anthropic- and OpenAI-style APIs.

50%
Unverified

Billion Context's cache hit rate for healthy sessions can reach 95% to 97%, and the compression itself costs 2% or less.

50%
Unverified

In the needle-in-a-haystack test, Billion Context's accuracy dropped from 69% to 27%.

50%
Unverified

Billion Context can be integrated with clients such as Claude Code, OpenCode, and Codex, and supports models such as DeepSeek V4 Flash, Claude 3.1 Pro, GLM-5.3, Qwen 3.8 Max, and Minimax M3.1 Flash.

50%
Unverified

In a 1.48-million-token test, full context failed all 199 tests, while Billion Context passed each time with a compression ratio of about 30x.

50%
Unverified

Billion Context can be installed via npm (npm install -g billion-context), uses the billy command, and supports three integration methods: installing as a native plugin, launching the client directly through the proxy, and manually prepending the proxy address to the base URL.

50%
Unverified

Billion Context is currently at version 0.1.21, auto-updates the main branch by default, has no authentication at all when run with the host parameter, and the desktop Code tab bypasses the proxy and is therefore not compressed.

50%
Unverified

The accuracy of needle-in-a-haystack exact retrieval fell from 69% with full context to 27% with Billion Context, because summaries retain core semantics rather than verbatim text.

50%
Unverified

Billion Context is a lightweight proxy between the Agent and the model API that compresses conversation history by pointing the Agent's base URL at it, and is compatible with both the Anthropic and OpenAI API formats.

50%
Unverified

Billion Context provides four tools for the model: compress (fold messages into summaries), decompress (retrieve the original text), search context (keyword search of compressed content), and status query (show context usage).

50%
Unverified

Billion Context's compression is incremental, summarizing only small-scope content, and the process is fully reversible, unlike built-in summarization that summarizes the whole text at once.

50%
Unverified

Officials say the prompt cache hit rate in normal conversations stays stable at 95% to 97%, and the overhead of compression itself is only 2% or even less.

50%
Unverified

Billion Context's official paper covers four and a half months of actual use, 174,000 model calls, and 18.76 billion input tokens, with zero context window overflow, all completed on 200k models.

50%
Unverified

In the Context Bench benchmark run on DeepSeek V4.1 Flash, Billion Context scored an F1 of 0.31, exceeding full context's 0.29; the cost was $4.61, 58% lower than full context's $11.05.

50%
Unverified

In the Long Memo Val benchmark, Billion Context scored 79.2%, full context 83.2%, and the best RAG approach 62.6%, with cost reduced by 71%.

50%
Unverified

All four of these plugins can be found in the DSH plugin marketplace by searching their names, or obtained by asking DSH directly.

50%

Source Articles