semble-hooks — automatic code context for every prompt via Semble semantic search

[benediktkraus]

What I built

semble-hooks — hooks that automatically inject the most relevant code chunks from your project into every prompt.

Powered by Semble semantic code search (98% token savings vs grep+read, CPU-only, 16MB static model, no API key).

How it works

You type a prompt
    ↓
SessionStart hook warms up the Semble index
    ↓  
UserPromptSubmit hook calls semble search with your prompt
    ↓
Top-5 code chunks injected as <relevant-code> into agent context
    ↓
Agent understands your codebase without grep/read cycles

Codex integration

• SessionStart → code-bootstrap.mjs (warms up index)
• UserPromptSubmit → code-recall.mjs (injects code chunks)
• Graceful degradation: if semble is not installed, hooks silently pass through
• Zero npm dependencies, pure Node.js built-ins

Install

pip install semble
git clone https://github.com/benediktkraus/semble-hooks.git
cd semble-hooks
./install.sh codex

Also supports Claude Code, Gemini CLI, and OpenClaw.

Why hooks instead of MCP?

Semble has an MCP mode, but with hooks every prompt gets code context automatically — no agent decision needed, no tool call overhead, no missed context.

Feedback welcome!