Neurarch MCP Server

neurarch-mcp

Model Context Protocol server that exposes a Neurarch model graph to Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, Codex, and any other MCP-aware AI agent.

The agent gets structural awareness of your neural network: layer list, parameter counts, FLOPs, blast-radius impact analysis, and Mermaid diagrams, without you pasting 200 lines of nn.Module into chat.

▶ Watch the 13s demo — every number above is produced by the tools, not guessed by the model.

Why

When you ask Claude Code "rewrite the training loop for my new encoder block", it sees your train.py but not your model. It guesses at shapes and parameter counts. neurarch-mcp plugs the structured model graph into the same conversation so the agent can answer "what depends on attn_8?" or "where does the parameter budget actually live?" with one call instead of ten.

What it looks like

A real exchange against examples/tiny-gpt.neurarch.json:

You ▸ Where does the parameter budget live, and is it safe to drop block_1?

Claude ▸ (calls describe_architecture)
  tiny-gpt — 8 layers, depth 7, 57.8M params, 8.27 GMACs.
  Heaviest by params: token_embedding (25.7M, 44%), lm_head (25.7M), block_0/1 (3.1M each).
  Heaviest by compute: lm_head, then the two decoder blocks. Validation: clean.

  (calls layer_impact { names: ["block_1"] })
  Downstream of block_1: final_norm and lm_head (shape-sensitive, weight-carrying), then output.

  The budget is dominated by the embedding + LM head (~89%), not the decoder blocks,
  so dropping block_1 frees only ~3.1M params (~5%) while removing one of just two
  decoder layers. It is structurally safe (final_norm just re-reads block_0's output),
  but it is the wrong lever if your goal is shrinking the model — quantizing the
  embedding/head would do far more.

Numbers above are produced by the tools, not estimated by the model.

Install

No install step. Every client below launches the server the same way:

npx -y neurarch-mcp /abs/path/to/your-model.neurarch.json

To produce the model file: open your model in the Neurarch app, then File → Save (.json). The MCP server reads that file directly. Add --watch so the agent sees app-side saves without a restart, and --write if you want the agent to be able to edit the model (off by default).

Use an absolute path to the model file in any global config: npx does not run from your project directory, so relative paths only work in project-scoped configs.

Claude Code

One command:

claude mcp add neurarch -- npx -y neurarch-mcp /abs/path/to/your-model.neurarch.json --watch

Or commit a project-scoped .mcp.json at the repo root so every collaborator gets the server automatically:

{
  "mcpServers": {
    "neurarch": {
      "command": "npx",
      "args": ["-y", "neurarch-mcp", "./model.neurarch.json", "--watch"]
    }
  }
}

Claude Desktop

Open Settings → Developer → Edit Config, or edit the file directly:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "neurarch": {
      "command": "npx",
      "args": ["-y", "neurarch-mcp", "/abs/path/to/your-model.neurarch.json", "--watch"]
    }
  }
}

Fully quit and reopen Claude Desktop (the config is read at startup). The tools appear under the search-and-tools icon in the chat input.

Cursor

Create .cursor/mcp.json in your project (or ~/.cursor/mcp.json for all projects), then enable the server under Settings → MCP:

{
  "mcpServers": {
    "neurarch": {
      "command": "npx",
      "args": ["-y", "neurarch-mcp", "./model.neurarch.json", "--watch"]
    }
  }
}

VS Code (Copilot agent mode)

Create .vscode/mcp.json (note the servers key, not mcpServers):

{
  "servers": {
    "neurarch": {
      "command": "npx",
      "args": ["-y", "neurarch-mcp", "${workspaceFolder}/model.neurarch.json", "--watch"]
    }
  }
}

Or from a shell: code --add-mcp '{"name":"neurarch","command":"npx","args":["-y","neurarch-mcp","/abs/path/to/model.neurarch.json"]}'

Other clients (Windsurf, Codex, ...)

Same command + args shape; only the config file location differs. For clients that speak Streamable HTTP instead of stdio, run the server with --http and point the client at it:

{
  "mcpServers": {
    "neurarch": {
      "type": "http",
      "url": "http://127.0.0.1:8787/mcp"
    }
  }
}

If you set NEURARCH_MCP_TOKEN, add "headers": { "Authorization": "Bearer <token>" }. See Remote access for tunnels and security.

Verify it works

Ask the agent: "List the Neurarch tools you can see." You should get describe_architecture, layer_impact, validate_model and friends (17 read tools; 6 more with --write). From a shell, npx -y neurarch-mcp --help prints usage and the full tool list.

Try it in 30 seconds (no app needed)

This repo ships runnable example models under examples/. Point the server at one and your agent can immediately answer structural questions:

{
  "mcpServers": {
    "neurarch": {
      "command": "npx",
      "args": ["-y", "neurarch-mcp", "./examples/tiny-gpt.neurarch.json"]
    }
  }
}

• examples/tiny-gpt.neurarch.json — a small GPT-style decoder (embedding, 2 transformer blocks, LM head).
• examples/tiny-cnn.neurarch.json — a CIFAR-style CNN (2 conv stages + classifier).
• examples/resnet-mini.neurarch.json — a residual block with a skip/merge node (a branchier graph for impact and path tools).

Then ask:

Look at the Neurarch model. Where do the parameters actually live, and which block would shrink the model fastest if I cut it in half?

The agent calls describe_architecture (one shot: pipeline, depth, param + compute hotspots, validation), then layer_impact on the heaviest block, and writes a recommendation grounded in the actual numbers from the model, like the transcript above.

Tools

Read (always available)

Write (opt in with --write)

layer_impact is the headline read tool. Before the agent recommends delete every conv_X, it can call layer_impact and tell the user "this rewires 8 downstream layers, 3 of which carry weights and will need rebuild." validate_model is the headline safety tool — call it before recommending a destructive edit to surface pre-existing issues separately from the change.

Flags

• --write — expose mutation tools. Off by default so accidental writes can't clobber a file you're editing in the Neurarch app.
• --watch — poll the model file for changes and reload on save. Pair with the Neurarch app: edit visually, agent sees the latest graph without restarting the MCP server. Note: an external save will overwrite any unsaved in-memory edits made via --write.
• --http[=PORT] — serve over Streamable HTTP instead of stdio (default port 8787). See Remote access below.
• --host=ADDR — bind address for --http. Defaults to 127.0.0.1 (loopback only).
• --version (alias -v) — print the version and exit. --help (-h) prints usage and the full tool list.

Remote access

By default the server talks stdio, so the agent and the model file live on the same machine. --http serves the same tools over Streamable HTTP, so a hosted or phone-based agent can drive a model running on your machine — e.g. behind a Cloudflare or Tailscale tunnel.

# local only (safe default: loopback, no auth needed)
npx neurarch-mcp model.neurarch.json --http

# expose to a tunnel with a bearer token and write tools
NEURARCH_MCP_TOKEN=$(openssl rand -hex 16) \
  npx neurarch-mcp model.neurarch.json --write --http --host=0.0.0.0
# then point cloudflared / tailscale funnel at :8787 and connect the agent to
# https://<tunnel>/mcp with the same token.

POST JSON-RPC to /mcp; GET /health is a liveness probe. Sessions follow the standard Streamable HTTP handshake (Mcp-Session-Id), so any MCP-aware client connects unchanged.

Security:

• Binds to 127.0.0.1 by default. Without a token, the Host header is checked against a loopback allowlist (DNS-rebinding protection) and no CORS headers are sent.
• Set NEURARCH_MCP_TOKEN to require Authorization: Bearer <token> on every request (constant-time checked). It is required before --write may bind to a non-loopback host — the server refuses to start otherwise.

Sharing results with the corpus (opt-in)

Set NEURARCH_REPORT=1 to share one anonymous structure+verdict row per validate_model call with the Neurarch corpus: the structural fingerprint (8-char hash), the layer-type histogram and edge count that let the server verify it, and the finding (rule id, severity) pairs. Never the graph, parameter values, layer names, file paths, or any identity: the payload shape cannot carry them, and the server rejects rows whose fingerprint does not recompute from the histogram it came with.

Off by default: without the flag this server makes no network calls at all. Reporting is fire-and-forget with a 5-second cap, so it can never slow or fail a tool call. Policy: neurarch.com/rules.html#data.

Troubleshooting

• The server never appears in the client. The model path must be absolute in any global config; npx does not run from your project directory. Relative paths only work in project-scoped configs (.mcp.json, .cursor/mcp.json, .vscode/mcp.json).
• Read tools work but write tools are missing. You did not pass --write. It is off by default so accidental writes can't clobber a file you're editing in the app.
• npx fails on first run. Node >= 20 is required (node --version).
• Claude Desktop shows nothing after editing the config. Fully quit and reopen the app; the config is only read at startup.
• The agent sees a stale graph after you edit in the app. Add --watch, or restart the server.

What this is not

• Not a generic codebase indexer. This serves one .neurarch.json file. For codebase structure, use GitNexus or similar.
• Not connected to your Neurarch workspace. v1 reads a saved JSON file only. Live editing happens in the Neurarch web app.

Issues & Feedback

This repo is the public home for both:

• neurarch-mcp (this MCP server): bugs, protocol changes, integration questions.
• Neurarch (the app): canvas bugs, agent issues, linter rules, feature requests.

Please tag issues with mcp, app, linter, or feature-request so we can triage faster.

Star this repo

If neurarch-mcp saved you from pasting an nn.Module into chat, a ⭐ helps other ML engineers find it. It is the lowest-effort way to support the project.

Contributing

A new tool is a small, self-contained PR. See CONTRIBUTING.md for the 3-step "add a tool" guide.

Development

git clone https://github.com/neurarch-ai/neurarch-mcp
cd neurarch-mcp
npm install
npm run typecheck             # tsc --noEmit
npm run build                 # tsup → dist/index.js
npm test                      # vitest (≈100 unit tests)
node dist/index.js --help     # confirm bin works

CI runs typecheck + build + test on Node 20 and 22 for every push and PR.

The package vendors a small set of pure-TypeScript utilities (model types, parameter and FLOP estimators, impact analyzer) from the main Neurarch repo. They live under src/lib/ and have no runtime dependencies beyond @modelcontextprotocol/sdk.

License

MIT. See LICENSE.

Links

• Neurarch — the visual neural-network editor that produces the model files this server reads.
• Model Context Protocol — the spec this server implements.
• npm — package page.