Fast local file discovery, source search, and smart block-clamped context for coding agents.
Runs as a native VS Code Language Model Tools extension (#cctxNativeContext, #cctxNativeGrep, #cctxNativeRead, #cctxNativeFiles) and MCP server backed by one shared Rust cache on Linux, native Windows (cctx.exe), WSL, and SSH, with <8 KiB inline response guards, rich document anchors (#toc, #slide-N-M, #heading), and a live model/token/cost dashboard (cctx usage).
A cold cache returns immediately so normal tools can continue.
Requirements: Linux, native Windows (cctx.exe, no WSL required), WSL, or Linux SSH; stable Rust; ripgrep (rg). The VS Code extension requires VS Code 1.140+ and automatically injects agent usage instructions (applyTo: "**").
1. Install the backend (Linux/WSL/SSH or native Windows PowerShell):
# Linux / WSL / SSH:
git clone https://github.com/dotnetpower/code-context.git
cd code-context && bash scripts/install.sh && ~/.local/bin/cctx ping
# Native Windows (PowerShell): cargo build --locked --release; copy target\release\cctx.exe $HOME\.local\bin\cctx.exe2. Install Code Context by dotnetpower from Marketplace,
or run code --install-extension dotnetpower.code-context-native (or install the VSIX).
The backend is installed separately; installing the extension alone is not enough.
Existing VSIX users should follow the publisher migration guide.
3. Open a trusted project and run Code Context: Getting Started (or cctx usage --hours 24 in the terminal for the live model, token, cache hit, and AIC/$ cost savings dashboard).
On native Windows, cctx.exe daemon starts automatically in the background on first use. Try:
Use
#cctxNativeFilesto find*Service*, then#cctxNativeContextto inspect a symbol.
No proposed API flags are needed. For Windows Agents, SSH, manual registration, or MCP-only clients, see the setup guide and backend reference. Marketplace and GitHub VSIX distribution use the same extension identity.
On AKS Store Demo, the fixed
mixed retrieval suite averaged 15.20× faster with a warm CCTX cache
(95% CI 14.53–15.84×, up from 6.83× in v0.14): 4.450 → 0.293 ms per operation, with a 29.115 ms cold sync median (down from 98.813 ms in v0.14).
Replacing multi-turn rg + view reads with one block-clamped code_context call cuts lookup payload tokens by 73.2–91.3%, rich document (#toc / #slide-N) reads by 86.3–98.9%, and end-to-end coding agent wall time by 24.7%.
| Retrieval task (AKS Store Demo) | CCTX off | CCTX on, warm | Off / on |
|---|---|---|---|
| Filename discovery | 4.923 ms | 0.307 ms | 16.03× |
| Literal search | 5.885 ms | 0.289 ms | 20.35× |
| Search + context | 5.935 ms | 0.317 ms | 18.70× |
| Known file read + hash | 0.083 ms | 0.251 ms | 0.33× — slower |
| Fixed mixed suite | 4.450 ms | 0.293 ms | 15.20× |
| Token & agent acceleration benchmark | Baseline (rg + view) |
With CCTX (code_context / code_read) |
Reduction / speedup |
|---|---|---|---|
| Rust 6-symbol lookup payload tokens | 15,028 tok (100L) / 25,133 tok (200L) | 4,031 tok (6 calls vs 20, 10.3× faster) | -73.2% to -84.0% tokens |
| Python 6-symbol lookup payload tokens | 13,997 tok (100L) / 24,609 tok (200L) | 2,141 tok (6 calls vs 15, 13.16× faster) | -84.7% to -91.3% tokens |
Rich documents (.pptx/.docx/.md #toc & #slide-N) |
Full-document read | Virtual #toc + #slide-N-M (5.4× faster DEFLATE) |
-86.3% to -98.9% tokens |
| Copilot A/B agent tool calls & LLM turns | 7.00 calls / 5.00 turns | 2.67 calls / 3.67 turns | -61.9% calls · -26.7% turns |
| Copilot A/B agent payload tokens & AIU cost | 6,360 tok / 4.76B nanoAIU | 3,177 tok / 3.55B nanoAIU | -50.0% tokens · -25.4% cost |
| Copilot A/B end-to-end task wall time | 20.31 s | 15.30 s (100% accuracy both) | -24.7% completion time |
Live multi-session telemetry (cctx usage --hours 24) |
Unmonitored | Per-model tokens, 96.6% warm hit & AIC/$USD savings | -10% to -39% session bill |
Eight daemon sessions, 40 rounds, 28 tasks, 8,960 correctness-checked pairs. Off bypasses CCTX with ripgrep/direct read/hash; on uses daemon RPC. Small known reads are faster without IPC. A cold miss plus fallback had a 4.837 ms median, not a warm-cache speedup. See method, cold-start chart, token suites, raw samples, and reproduction for uncertainty and limitations.