Don’t take our word for it. The self-benchmarking suite proves the 40–70× claim on a committed fixture in CI — but your codebase isn’t our fixture. The only way to know what NeuralMind does for you is to run it on your code.
This walkthrough gets you from zero to a real before/after number on your repository in under 5 minutes, with no commitment beyond a pip install.
If the numbers don’t justify the install — you uninstall and move on. Nothing else happens. NeuralMind never uploads anything.
pip install neuralmind tiktoken
graphifyy builds the code knowledge graph NeuralMind reads from; tiktoken is needed for accurate OpenAI-model token counting. If you only care about a rough number, tiktoken is optional.
cd /path/to/your-project
neuralmind build .
The first build takes 1–3 minutes depending on repo size. Incremental rebuilds after code changes take seconds.
Sanity check — make sure it worked:
neuralmind stats .
You should see something like:
Project: your-project
Built: True
Nodes: 1,247
Communities: 23
If Built: False, the build step failed — see Troubleshooting.
neuralmind benchmark .
Output (your numbers will vary):
Project: your-project
Wake-up tokens: 412
Avg query tokens: 891
Avg reduction: 46.0x
Summary: NeuralMind query returns 46x less context than loading files naively
What those numbers mean for you:
| Metric | What it says |
|---|---|
| Wake-up tokens | Cost of one “orient the agent” call at session start. ~400 tokens = ~$0.0012 on Claude Sonnet. |
| Avg query tokens | Cost of one code question (across NeuralMind’s default 5-query sample). ~900 tokens = ~$0.0027 per question. |
| Avg reduction | How many times smaller NeuralMind’s context is vs loading whole files. 46× means your bill drops by ~97.8% per query. |
At 100 queries/day on Claude 3.5 Sonnet ($3/MTok input):
Adjust for your model and volume — the math scales linearly. GPT-4o costs ~5× more than Sonnet, so savings are larger. Claude Haiku costs less, so savings are smaller in absolute terms but the ratio stays the same.
You now have measured, reproducible numbers on your code, not ours. Three paths:
If the savings justify it, nothing more to do — you’ve already installed. Start asking code questions:
neuralmind query . "How does authentication work?"
neuralmind skeleton src/auth/handlers.py
Claude Code users: install the PostToolUse compression hooks for an extra 5–10× reduction layer on top:
neuralmind install-hooks .
neuralmind benchmark . --contribute --submitter your-github-handle
That flag emits a JSON blob with your project name, numbers, and the exact command that produced them. Nothing is uploaded. You get a text blob to paste into Slack, a design doc, or a PR.
If your project is open source (or the numbers are OK to share), drop the JSON blob into the community leaderboard:
--contribute output.docs/community-benchmarks.json and run python scripts/render_community_table.py --inject README.md.Every submission is auditable — entries include the exact neuralmind benchmark command that produced them.
A few things to check before giving up:
neuralmind stats . should report a non-zero node count. If it’s tiny, graphify may have missed your language or the project structure.sample_queries if you use the Python API.