A semantic context pre-processor that gives coding agents only the code they need.

Prune your codebase to only what's relevant before the LLM sees it.

Inspired by MiniMax M3's MSA architecture — sparse attention with pre-filtering.
Instead of attending to everything, identify which blocks matter and drop the rest.

ContextSlice is a sparse context compiler for AI coding agents.

• Before: Your coding agent sees 80 files and guesses what matters.

• After: ctxslice emits 12 relevant files/chunks under a 40K token budget.

Cursor dumps your whole codebase into context. Most of it is noise for any given task.

ctxslice runs before you open Cursor. It:

1. Chunks your codebase at AST level (function/class granularity, not file-level)
2. Embeds chunks into a persistent local vector index
3. Scores each chunk across three signals:
   • Semantic similarity to your task description
   • Whether the chunk is in the import graph of your active file
   • Whether the chunk overlaps with your current git diff
4. Packs the highest-scoring chunks into a token budget
5. Outputs a .ctxslice.md file — drag it into Cursor as @context

pip install -e .

ctxslice run "add pagination to UserList"

ctxslice run "add pagination to UserList" \
  --file src/components/UserList.tsx \
  --tokens 40000

ctxslice run "fix auth token refresh" \
  --file src/auth/TokenService.java \
  --output context/auth-task.md

ctxslice index

ctxslice stats

ctxslice index --force

ctxslice clear

Full documentation — installation, step-by-step usage with Cursor, Claude, and GitHub Copilot, CLI reference, and troubleshooting: ContextSlice-UserDocumentation.md

Release notes — see RELEASE_NOTES.md for changes by version.

1. Run ctxslice run "your task" --file your-active-file.ts
2. Open Cursor
3. In the chat, type @ and select .ctxslice.md
4. Ask your question — Cursor now has only the relevant parts of your codebase

Or add to .cursorrules:

Always read @.ctxslice.md for context on the current task.

Composite score thresholds:

• = 0.60 -> HIGH — included, full content

• = 0.35 -> MEDIUM — included, full content

• < 0.35 -> DROPPED — not emitted

Files in the import graph that didn't make either tier appear in the DEPENDENCY SURFACE section (collapsed, for type/interface correctness).

ctxslice is written in Python but works on any language codebase — the language of your project has nothing to do with the tool's runtime.

AST-level chunking (function/class granularity) is supported for:

TypeScript, JavaScript, Java, Python, Go, Rust, Kotlin, Swift, C/C++, C#, Ruby

Files with other extensions are not skipped — they fall back to whole-file chunks and still contribute to context, just without method-level splitting.

• First run: index build takes 1–3 minutes for a large repo (embedding model downloads ~22MB)
• Subsequent runs: incremental — only changed files are re-embedded, typically 2–5s
• Deleted files are automatically purged from the index on each run; no manual cleanup needed
• Index stored at <repo>/.ctxslice/index/ — add to .gitignore
• --tokens minimum is 1000; default is 40000

.ctxslice/
.ctxslice.md

ctxslice/
├── chunker.py    # AST-level splitting via tree-sitter
├── index.py      # Persistent ChromaDB vector index + sentence-transformers
├── signals.py    # Git diff signal + import graph signal
├── scorer.py     # Composite scoring + token-budget packing
├── renderer.py   # .ctxslice.md output formatter
└── cli.py        # Typer CLI (run / index / stats / clear)

If you find ContextSlice useful, you can support the project:

👉 https://buymeacoffee.com/nishchya

It helps keep the project alive and growing.

MIT — see LICENSE.