> ## Documentation Index
> Fetch the complete documentation index at: https://docs.neuraldraft.io/llms.txt
> Use this file to discover all available pages before exploring further.
# MCP setup
> Install `@neuraldraft/mcp` in Claude Code, Cursor, Continue, or any other MCP-capable client.
`@neuraldraft/mcp` is a stdio MCP server published on npm. Every supported
client spawns it as a local process and talks JSON-RPC over stdio. You only
need three things:
1. Node 18.18 or newer.
2. A Neural Draft API key (`ndsk_live_...` or `ndsk_test_...`).
3. Your editor of choice.
For best results, mint a separate API key per machine — your audit log will
show "claude-code-laptop", "cursor-mac-mini" and so on rather than a single
ambiguous "personal".
## Claude Code
The fastest path. The CLI handles everything:
```bash theme={null}
claude mcp add neural-draft \
--env NEURALDRAFT_API_KEY=ndsk_live_... \
-- npx -y @neuraldraft/mcp
```
That writes the server config to `~/.config/claude-code/mcp.json` (or your
platform's equivalent). To verify, restart Claude Code and run `/mcp` — you
should see `neural-draft` listed under "Connected servers".
If you'd rather edit the config by hand:
```json theme={null}
{
"mcpServers": {
"neural-draft": {
"command": "npx",
"args": ["-y", "@neuraldraft/mcp"],
"env": {
"NEURALDRAFT_API_KEY": "ndsk_live_..."
}
}
}
}
```
## Cursor
Cursor reads `.cursor/mcp.json` either at the workspace root (project-scoped)
or at `~/.cursor/mcp.json` (global).
```json theme={null}
{
"mcpServers": {
"neural-draft": {
"command": "npx",
"args": ["-y", "@neuraldraft/mcp"],
"env": {
"NEURALDRAFT_API_KEY": "ndsk_live_..."
}
}
}
}
```
Reload the workspace. Open the AI panel and confirm `neural-draft` appears
under "MCP". When you ask Cursor to generate a section, it'll auto-call
`register_component`.
Putting an API key in a workspace file checked into git is a footgun. Use
the global `~/.cursor/mcp.json` for personal keys, or a `.cursor/mcp.json`
with an env var indirection (e.g. `"NEURALDRAFT_API_KEY": "${env:NEURALDRAFT_API_KEY}"`)
and `.gitignore` the file.
## Continue
Add the server to `~/.continue/config.yaml`:
```yaml theme={null}
experimental:
modelContextProtocolServers:
- transport:
type: stdio
command: npx
args: ["-y", "@neuraldraft/mcp"]
env:
NEURALDRAFT_API_KEY: ndsk_live_...
```
Reload Continue. The Neural Draft tools become available to any model you've
configured.
## Lovable, v0, Bolt — the workaround
These hosted tools don't yet support MCP. For now, paste the system prompt at
[frameworks/lovable](/frameworks/lovable) into your build prompt — it teaches the
model the same conventions the MCP server enforces. When Lovable / v0 / Bolt
add MCP support, this server will work as-is.
## Multiple projects on one machine
The cleanest pattern is **one server entry per project**, named for the
project. The AI client UI shows them separately, so you can flip between
projects without changing env vars:
```json theme={null}
{
"mcpServers": {
"neural-draft-acme": {
"command": "npx",
"args": ["-y", "@neuraldraft/mcp"],
"env": { "NEURALDRAFT_API_KEY": "ndsk_live_acme..." }
},
"neural-draft-bigcorp": {
"command": "npx",
"args": ["-y", "@neuraldraft/mcp"],
"env": { "NEURALDRAFT_API_KEY": "ndsk_live_bigcorp..." }
}
}
}
```
When asking the AI to "use Neural Draft", reference the entry by name — e.g.
"use the neural-draft-acme tools". Most clients pick automatically based on
which server has the relevant tool, but explicit is better for clarity.
## Environment variables
The server reads these env vars (the `NEURAL_DRAFT_*` aliases are also
accepted for backwards compatibility):
| Variable | Required | Default |
| -------------------------- | -------- | ------------------------------- |
| `NEURALDRAFT_API_KEY` | Yes | — |
| `NEURALDRAFT_API_URL` | No | `https://api.neuraldraft.io/v1` |
| `NEURALDRAFT_PROJECT_ID` | No | (resolved from the API key) |
| `NEURALDRAFT_DISPLAY_NAME` | No | (auto-generated) |
Override `NEURALDRAFT_API_URL` only when pointing at a self-hosted API or our
staging environment (`https://api.staging.neuraldraft.io/v1`).
## Verifying it works
After install, ask the AI:
> *"Read the Neural Draft brand context for this project."*
You should see the AI invoke the `brand://current` resource. If it doesn't,
the server didn't connect. Common causes:
| Symptom | Fix | |
| ---------------------------------------- | ------------------------------------------------------------------------------------------ | ------------------------------------------- |
| `npx: command not found` | Install Node 18.18+. `node --version`. | |
| Server starts then crashes immediately | Check the API key prefix; it must match \`ndsk\_(live | test)\_...\`. The server validates on boot. |
| AI doesn't see the tools | Most clients require a workspace reload after editing the MCP config. Restart the editor. | |
| Cursor shows "neural-draft" but no tools | Stale Node cache. `npx -y @neuraldraft/mcp@latest` to force re-download. | |
| `429` on every tool call | You're inside a tight loop. Throttle your AI's session, or upgrade to a higher rate limit. | |
If none of those: run the server manually with the inspector to see the raw
JSON-RPC transcript:
```bash theme={null}
NEURALDRAFT_API_KEY=ndsk_test_... npx @modelcontextprotocol/inspector \
npx -y @neuraldraft/mcp
```
Open `http://localhost:6274` and try `brand://current` directly.
## What you get next
The full reference: every tool, every resource, every parameter.
Drop-in framework integrations once you've got the AI generating Neural
Draft-native code.