ContextCore

Search all your local data — notes, code, recordings, images — and send only what matters to AI.

Reduce LLM context tokens by ~57% on SciFact benchmark settings.

Install

Install from PyPI:

python -m pip install contextcore==1.0.0

Optional source install (for contributors):

git clone https://github.com/lucifer-ux/SearchEmbedSDK.git
cd SearchEmbedSDK
python -m pip install -e .

Then run the setup wizard:

contextcore init

Gif is sped up to skip the installation parts. This is not supermemory, but supercharged memory for all file formats shared all across.

That's it. ContextCore indexes your files, registers with your AI tools, and runs in the background. No config files to edit.

Prerequisites

• Python 3.10+
• Windows, macOS, or Linux
• Internet access for first-time model downloads
• Enough disk space for Python packages and model files

Optional but important:

• ffmpeg for video indexing
• Claude Desktop or another MCP-capable tool if you want interactive AI integration

What ContextCore Does

ContextCore gives you:

• a CLI command: contextcore
• a local backend server, normally on http://127.0.0.1:8000
• an MCP server script for Claude and similar tools
• local indexing for:
  • text and documents
  • images
  • audio transcripts
  • video embeddings and video context
  • codebase context (structure, symbols, dependencies, file-level detail)

Codebase Context for Claude/OpenCode

ContextCore can expose your codebase context directly to MCP tools (for example, Claude Desktop and OpenCode) so the model can reason over your project without you pasting the entire directory into chat.

Use the code modality during setup (contextcore init) and ContextCore will provide indexed codebase context through MCP tools such as:

• get_codebase_context
• get_codebase_index
• get_module_detail
• get_file_content

Recommended Setup

For real usage, the most reliable setup is:

• keep one dedicated Python virtual environment
• use that same Python environment for:
  • contextcore init
  • contextcore serve
  • mcp_server.py in your Claude config

Do not test the backend in one venv and point Claude at a different venv. That is one of the most common causes of "it works in the terminal but not in Claude".

Verify Install

Run:

contextcore --help

If that fails, the package is not installed in the Python environment your shell is using.

BEIR Benchmark (SciFact)

You can benchmark the current text retrieval stack on a BEIR dataset (starting with SciFact) without touching your existing index data.

Install optional benchmark dependency:

python -m pip install beir

Run benchmark:

contextcore benchmark --dataset scifact --top-k 10

Optional fast iteration with fewer queries:

contextcore benchmark --dataset scifact --top-k 10 --max-queries 50

Optional JSON output:

contextcore benchmark --dataset scifact --output-json .\benchmarks\scifact_run.json

Token reduction benchmark (tiktoken):

python -m pip install tiktoken
contextcore benchmark --dataset scifact --top-k 10 --measure-tokens --context-top-k 5

Compare retrieval systems (ContextCore vs BM25) and export publish-ready tables:

contextcore benchmark --dataset scifact --top-k 10 --measure-tokens --context-top-k 5 --systems contextcore_hybrid,bm25_only,trigram_only --report-csv .\benchmarks\scifact_compare.csv --report-md .\benchmarks\scifact_compare.md --output-json .\benchmarks\scifact_compare.json

Daily C

... View full README on GitHub