@neuronsearchlab/mcp

MCP (Model Context Protocol) server for NeuronSearchLab. Gives any MCP-compatible AI client (Claude, Codex, Cursor, Windsurf, etc.) direct access to NeuronSearchLab recommendations, product/content search, catalogue operations, analytics, and platform controls.

"Get 5 recommendations for user alice@example.com"
"Check my current plan and show which resources are over their limits"
"Create a new context called Twitter Feed"
"Add a pin rule so Nike items always appear in the top 3"
"Why did item prod-456 rank first for bob?"

Two ways to run it:

• Hosted (recommended, no install): https://console.neuronsearchlab.com/api/mcp — Streamable HTTP with OAuth sign-in or an NSL API key. Listed on the MCP Registry as com.neuronsearchlab/mcp (search neuronsearchlab in the registry API or directory).

• Local stdio via npm: npx -y @neuronsearchlab/mcp in two modes — public (recommendations, events, catalogue via OAuth client credentials) or internal (admin platform via console API key).

Connect to the hosted server (no install)

The hosted endpoint runs a submission-safe customer administration profile. It includes first-class tools for plan and usage visibility, ranking configuration, experiments, training, analytics, catalogue inspection, API-key inventory and revocation, integrations, and event types. Every hosted tool declares its OAuth requirement and requires the authenticated team's admin scope. Credential creation, billing mutations, and the arbitrary platform API fallback remain available only to trusted local/internal clients so secrets, purchases, and unbounded API calls are not exposed in ChatGPT. Keys minted through OAuth consent appear in console → Security and can be revoked there anytime.

claude.ai / Claude Desktop — Settings → Connectors → Add custom connector → paste https://console.neuronsearchlab.com/api/mcp → Connect, then sign in to your NeuronSearchLab console and approve the scopes.

Claude Code

# OAuth (browser sign-in):
claude mcp add --transport http neuronsearchlab https://console.neuronsearchlab.com/api/mcp
# …or with an API key:
claude mcp add --transport http neuronsearchlab https://console.neuronsearchlab.com/api/mcp \
  --header "Authorization: Bearer nsl_your_key"

OpenAI Codex — in ~/.codex/config.toml:

[mcp_servers.neuronsearchlab]
url = "https://console.neuronsearchlab.com/api/mcp"
bearer_token_env_var = "NSL_API_KEY"

Cursor / Windsurf / other Streamable HTTP clients

{
  "mcpServers": {
    "neuronsearchlab": {
      "url": "https://console.neuronsearchlab.com/api/mcp",
      "headers": { "Authorization": "Bearer nsl_your_key" }
    }
  }
}

Related MCP server: FHIR MCP Server

Resources

• Product site: https://www.neuronsearchlab.com

• MCP setup guide: https://docs.neuronsearchlab.com/sdk/mcp

• AI agents for recommendation operations: https://www.neuronsearchlab.com/blog/ai-agents-for-recommendation-operations

• Recommendation systems reading path: https://www.neuronsearchlab.com/blog/recommendation-systems

Tools

API tools

Modes

Public mode

Uses OAuth client credentials and the public API.

Supported:

• recommendations

• events

• catalogue operations

Internal mode

Uses a NeuronSearchLab API key with the admin scope against the console API.

Currently supported:

• account plan and limits: get_account_plan returns the effective plan, resolved limits, current usage, and exact overages

• catalogue search and ranking debug: search_items, explain_ranking

• contexts: list_contexts, create_context, update_context, delete_context, get_context

• rules: list_rules, create_rule, update_rule, delete_rule, toggle_rule, enable_rule, disable_rule, get_rule

• segments: list_segments, get_segment, create_segment, update_segment, delete_segment

• experiments: list_experiments, get_experiment, create_experiment, update_experiment, start_experiment, stop_experiment, get_experiment_results, refresh_experiment_results

• training: list_training_jobs, get_training_job, create_training_job, cancel_training_job

• analytics: get_ranking_metrics, get_user_analytics, get_item_analytics, compare_items, top_items

• event types: list_event_types, create_event_type, update_event_type, delete_event_type

• credentials and integrations: list_api_keys, revoke_api_key, list_integrations (create_api_key is local/internal only because it returns credential material)

• fallback UI coverage for trusted local/internal clients only: list_platform_routes, call_platform_api

Quickstart (local stdio)

1. Get credentials

Generate SDK Credentials (OAuth 2.0 client ID + secret) from the NeuronSearchLab console.

2. Add to Claude Desktop

Public mode (recommendations, events, catalogue):

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "neuronsearchlab": {
      "command": "npx",
      "args": ["-y", "@neuronsearchlab/mcp"],
      "env": {
        "NSL_CLIENT_ID": "your-client-id",
        "NSL_CLIENT_SECRET": "your-client-secret"
      }
    }
  }
}

Internal mode (admin platform — contexts, rules, analytics, etc.):

{
  "mcpServers": {
    "neuronsearchlab": {
      "command": "npx",
      "args": ["-y", "@neuronsearchlab/mcp"],
      "env": {
        "NSL_PLATFORM_MODE": "internal",
        "NSL_API_KEY": "your-admin-api-key"
      }
    }
  }
}

Restart Claude Desktop. You'll see a 🔌 neuronsearchlab indicator in the toolbar when it's connected.

Try it: recommendation-ops demo

Once connected, run this named demo path before wiring the server into a larger workflow. It proves that an AI client can operate the recommender as an investigation surface rather than just call a recommendation endpoint.

1. List contexts so the assistant confirms the exact surface it is about to inspect.

2. Fetch recommendations for a known test user and context.

3. Search the catalogue for a concrete product/content phrase and compare the returned item IDs with the recommendation set.

4. Explain one ranked item using the request_id from the recommendation or search response when available.

5. Optional, sandbox only: draft or toggle a rule after the explanation shows the expected leak. Keep production contexts read-only unless the operator explicitly approves a write.

Use the NeuronSearchLab MCP server in read-only mode first.
List my recommendation contexts and choose the homepage-feed context.
Get 10 recommendations for user demo-user@example.com using context homepage-feed.
Search the catalogue for waterproof jackets and show the top 5 item ids.
Explain why the first recommended item ranked first, using the request_id from the recommendation response if available.
If you see a relevance leak, draft the smallest rule that would fix it, but do not create or toggle the rule yet.

Internal/admin mode can also inspect operational state before making changes:

List ranking rules for the homepage-feed context.
Show the latest ranking metrics for that context.
Compare item jacket-123 with item jacket-456 and explain which rule or signal separates them.

Next steps after the smoke test:

• create a scoped API key for the client or MCP server

• connect one real recommendation context, such as homepage-feed

• add request attribution to click/view events before judging ranking quality

3. Cursor / other MCP clients

Follow your client's MCP server guide. The command is:

npx @neuronsearchlab/mcp

Set NSL_CLIENT_ID + NSL_CLIENT_SECRET for public mode, or NSL_PLATFORM_MODE=internal + NSL_API_KEY for internal mode.

Releases

This repo uses Changesets plus GitHub Actions for automated versioning and npm publishing.

• Add a changeset for any user-facing package change with npm run changeset

• Merge that PR into main

• The release.yml workflow opens or updates a version PR

• Merging the version PR publishes @neuronsearchlab/mcp to npm automatically

To enable trusted publishing, configure the package on npmjs.com to trust the release.yml workflow in this repository.

Configuration

All configuration is via environment variables:

Tool reference

get_recommendations

Fetch personalised recommendations for a user. Returns ranked items with scores and a request_id for attribution.

Inputs

Example

Get 10 recommendations for user alice@example.com using context 101

get_recommendation_sets

Fetch every recommendation rail of a page in one request (POST /v1/recommendations/batch). Each set is served and billed on its own and carries its own request_id for attribution; a set that fails reports its own error. Not exposed in the ChatGPT app profile.

Inputs

Example

Get the homepage rails for alice@example.com: 12 from context 101 as "for_you" and 8 from context 102 as "trending"

get_auto_recommendations

Fetch the next auto-generated section for a user's feed. Designed for infinite-scroll — each call returns one curated section (e.g. "Trending this week", "New for you") plus a cursor for the next section. Call until done: true.

Inputs

track_event

Record a user interaction. Always pass request_id from the recommendations response to enable click-through attribution.

Inputs

upsert_item

Add or update an item in the catalogue. The description field is used to generate the embedding — write it to be rich and descriptive.

Inputs

The item ID is generated by NSL and returned in the ingestion response; callers cannot supply it.

patch_item

Partially update an existing catalogue item.

Inputs

delete_items

Permanently remove items. Cannot be undone. To temporarily exclude, use patch_item with active: false.

Inputs

search_items

Search the catalogue by keyword.

Inputs

explain_ranking

Explain why a specific item was ranked at a given position for a user. Returns score breakdown, applied rules, and pipeline trace.

Inputs

list_contexts

List all recommendation contexts (feeds) configured for your team.

Inputs — none

create_context

Create a recommendation context: one rail or surface. A context owns its model and its ranking configuration.

Inputs

Example

Create a new context called "Homepage feed"

get_context

Get one context: its fields, its model, its ranking configuration (template and serving stages) and configuration_enabled (whether serving uses that configuration).

update_context

Update a context. Only the fields you pass change.

Inputs

model is one of:

• { "kind": "automatic" }: the NSL default model.

• { "kind": "nsl", "family": "watch_next", "architecture": "gsasrec" }: an NSL model deployed to production.

• { "kind": "trained", "id": 41 }: a trained model. It supplies its own candidates, so candidate sources and learned reranking do not apply. Personalized discovery and email digest contexts only.

delete_context

Permanently delete a context, its ranking configuration and its rules.

Inputs

list_rules

List ranking rules, optionally filtered by context_id.

create_rule

Create a ranking rule. Rule types:

Inputs

Example

Create a pin rule called "Pin Nike" that pins items where brand equals "Nike" to position 3, scoped to context 1

update_rule / delete_rule / toggle_rule / enable_rule / disable_rule

Update, delete, or enable/disable a rule by rule_id.

get_user_analytics

Get served counts, event breakdown, unique-item activity, and click-through rate for a specific user.

Inputs

get_item_analytics

Get served counts, event breakdown, watch/click counts, and click-through rate for a specific item.

Inputs

compare_items

Compare two items head-to-head by served count, events, clicks, and CTR over the same time window.

Inputs

top_items

List the top items by served count or by matching event activity over a time window. Use metric="served" for generic "top item" or "best performing" questions. Use metric="events" when the user explicitly names an engagement signal (e.g. watch, click, purchase).

Inputs

Example

What's the top item served in the last 7 days?
Which items had the most watch events last month?

Authentication

Public mode uses OAuth 2.0 Client Credentials. Tokens are fetched on startup, cached in memory, and auto-refreshed 60 seconds before expiry.

Internal mode uses a NeuronSearchLab API key with the admin scope. Set NSL_API_KEY and NSL_PLATFORM_MODE=internal.

Development

git clone https://github.com/NeuronSearchLab/mcp
cd mcp
npm install
export NSL_CLIENT_ID=your-client-id
export NSL_CLIENT_SECRET=your-client-secret
npm run dev           # dev mode (tsx, no build)
npm run build         # compile to dist/

License

MIT