NeuroMesh
Your AI agent does not need three thousand lines of boilerplate to fix one bug. NeuroMesh indexes your repo into a live graph, routes your prompt to the right symbols, and sends a folded evidence packet — not whole-file dumps. Default engine: fast needs no model download.
Works with · Cursor · VS Code · Claude · Codex · OpenCode · Windsurf · Zed · Kilo · Trae · Antigravity · MiMo CLI
Your editor is oversharing
You ask one focused question. The agent pulls in two or three massive files, burns tokens, and still misses the function you cared about.
You pay for noise
Private helpers, legacy parsers, and unrelated utilities — billed on every turn.
You wait for the wrong stuff
The chat fills up while the model still has not seen the code path you asked about.
The model hallucinates
Too much context. Wrong conclusions. Bugs that were never in your codebase.
| What people try | Why it fails |
|---|---|
| Vector RAG | Chunks break functions apart. Structure disappears. |
| "Attach the whole file" | The model sees everything. Understands nothing. |
| Static code graphs | Nice map — then they still paste full files into the prompt. |
| NeuroMesh | Route, then fold. Only the relevant skeleton reaches the model. Expand a fold when you need the body. |
Keep structure. Cut tokens.
NeuroMesh never deletes your code. It folds what the task does not need into a single reversible line — like a table of contents that still knows where every function lives.
What you need — open
The functions and files that match your task stay fully visible with real bodies.
What you do not — folded
Everything else becomes a one-line marker with a fold ID. Signatures and imports stay in view.
Need the body? One tool call
Run neuromesh_expand_fold — instant restore from memory, no grep hunt.
Inspired by how nature packs information
Smart folding
Hide boilerplate. Keep signatures, imports, and file shape.
Shortest path
Connect only the files your task needs — not the whole repo neighborhood.
Remembers what worked
Files you actually changed get priority on the next similar task.
Safety modes
Balanced by default. Auth and payment tasks automatically get more context.
Faster follow-ups
The next likely file is warmed before your second tool call.
Live code graph
Functions, imports, and calls — so routing is structural, not guesswork.
3D Neural Galaxy
Run neuromesh monitor and explore your project as a live graph — subsystems, files, symbols — before you ask the agent anything.
Default URL: http://127.0.0.1:8765
· Port: neuromesh port
One question. One tight packet.
Ask in plain language (English, Persian, Arabic, …). NeuroMesh finds symbols, pulls neighbors, folds the rest, and tells you when to grep for more.
Prompt
Read the task exactly as written — intent is not lowercased into mush.
max_savings
Tiny, obvious edits
balanced
Everyday development
max_quality
Refactors, auth, critical paths
The more you ship, the better it routes
After a successful edit, call neuromesh_record_feedback. NeuroMesh strengthens the paths you actually used — so the next similar task needs fewer greps and fewer tokens.
Learns from real edits
Only files you actually changed influence the next packet — no fake shortcuts.
Shared paths stick
Files you often edit together stay linked — the graph prefers routes that worked before.
Ready for the next hop
Likely follow-up files are pre-warmed so expand and trace feel instant.
Measured on a real monorepo
Token savings are per task, after folding — not a marketing average. Run neuromesh eval on your repo.
v0.9.11 · default fast engine · up to ~97% fewer tokens vs attaching the whole project
| Task | Mode | Full repo | Before fold | Packet | Saved | Grep | ms |
|---|---|---|---|---|---|---|---|
| Fix MCP tool handler | balanced | 650,859 | 72,428 | 17,389 | 97.3% | 0 | 22 |
| Trace graph routing | balanced | 650,859 | 19,625 | 4,080 | 99.4% | 0 | 12 |
Fast by default. Embeddings opt-in.
v0.9.11 ships with engine: fast — instant graph index, no ONNX warm. Install, index, pass your prompt — done. Opt in to hybrid / deep when you need semantic search.
Retrieval engine
fast (default) — graph + prompt expansion, no download. hybrid / deep — add MiniLM when NL recall needs a boost.
neuromesh config engine get neuromesh install embed minilm # only for hybrid/deep neuromesh config engine hybrid
Connect your editor
One command registers MCP and copies agent rules so your IDE actually uses NeuroMesh instead of raw file reads.
neuromesh connect --global --agent-rules neuromesh index
Install NeuroMesh
One command, then neuromesh doctor to verify, and neuromesh connect to wire your editor.
curl -fsSL https://raw.githubusercontent.com/pinoox/neuromesh/main/install.sh | bash
irm https://raw.githubusercontent.com/pinoox/neuromesh/main/install.ps1 | iex
git clone https://github.com/pinoox/neuromesh.git cd neuromesh cargo build --release --bin neuromesh --features embeddings
cargo install --git https://github.com/pinoox/neuromesh.git neuromesh-cli --bin neuromesh --features embeddings
neuromesh doctor neuromesh connect --global --agent-rules neuromesh index # graph in RAM — usually under 30s
Connect in one command
NeuroMesh speaks MCP over stdio — the standard protocol Cursor, Claude, VS Code, Windsurf, Zed, and dozens of agents already support. No cloud. No API key for indexing. Your code stays on your machine.
{
"mcpServers": {
"neuromesh": {
"command": "neuromesh",
"args": ["mcp"]
}
}
}
Run neuromesh connect and pick your apps — or neuromesh connect --print to copy a snippet manually.
What your agent can call
Teach the agent this loop: get_context_packet → expand a fold if needed → trace callers → record_feedback after a good edit.
30+ languages and frameworks — Rust, TypeScript, Python, Go, PHP, Laravel, Django, Next, Vue, Spring, Android, Express, and more.