Skip to main content

​
Overview

OGX is an open, production-ready agentic API server and the successor to Llama Stack. It runs the agentic loop (inference, tool calling, RAG, and conversation management) behind the native APIs of OpenAI, Anthropic, and Google. Nimble ships as the remote::nimble-search provider, registered under the builtin::websearch toolgroup alongside the other web search backends. When a model calls the web_search tool, OGX routes the request to Nimble’s Search API and returns ranked, grounded results with source URLs. No fork, no plugin. The provider ships in OGX v1.1.0 and later.

​
Web search for any model

Any model OGX serves can search the web through Nimble, with no model-specific code. A web_search tool call resolves to builtin::websearch, which dispatches to the registered remote::nimble-search provider automatically. Nimble handles anti-bot evasion, JavaScript rendering, and geo-targeting underneath, so the model gets back clean, ranked results to ground its answer in.

​
Quick Start

​
1. Get a Nimble API key

Get your API key from Nimble’s dashboard (free trial available) and export it in the environment OGX runs in: 
export NIMBLE_API_KEY="your-api-key"

​
2. Register the provider and start OGX

Add Nimble to the tool_runtime section of your distribution config (for example, the starter distribution): 
tool_runtime:
  - provider_id: nimble-search
    provider_type: remote::nimble-search
    config:
      api_key: ${env.NIMBLE_API_KEY:=}
      max_results: 3
      search_depth: lite
      country: US
      locale: en
Then start the server: 
ogx run starter
On boot, OGX auto-registers the toolgroup: provider_id=nimble-search toolgroup_id=builtin::websearch.

​
3. Search the web from a model

Call the Responses API with the web_search tool enabled. OGX routes the call through builtin::websearch to Nimble: 
curl http://localhost:8321/v1/responses \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-4o-mini",
    "input": "What were the biggest AI model releases this month? Cite your sources.",
    "tools": [{"type": "web_search"}],
    "stream": false
  }'
The response contains a web_search_call item (status completed), followed by an assistant message grounded in live Nimble results with source URLs.

​
Configuration

The remote::nimble-search provider accepts these keys under config:
Required. Your Nimble API key. Provide it once via the NIMBLE_API_KEY environment variable (referenced as ${env.NIMBLE_API_KEY:=}), or per request via the X-OGX-Provider-Data header:
{"nimble_search_api_key": "your-api-key"}
A per-request header value overrides the configured key.
Optional. Maximum number of results to return. Defaults to 3.
Optional. Controls the speed-vs-richness tradeoff. Defaults to lite.
  • lite: titles, URLs, and snippets. Lowest latency, best for broad discovery.
  • deep: full real-time page extraction for each result. Higher latency, richer content, and may require a higher Nimble tier. If your account lacks it, the provider returns a clear “tier not enabled” message rather than failing.
See Search Depth for details.
Optional. ISO-2 country code for geo-targeted results (e.g. US, GB). Defaults to US. Can also be set per request via the model’s location hint.
Optional. Language/locale code for results (e.g. en). Defaults to en.

​
Using Nimble over MCP (alternative)

On an OGX release earlier than v1.1.0, or if you prefer an MCP-based setup, OGX can reach Nimble through its built-in Model Context Protocol support, with no provider configuration needed. Point a Responses mcp tool at Nimble’s hosted MCP server: 
curl http://localhost:8321/v1/responses \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-4o-mini",
    "input": "What is the NVIDIA AI-Q blueprint? Search the web and cite sources.",
    "tools": [{
      "type": "mcp",
      "server_label": "nimble",
      "server_url": "https://mcp.nimbleway.com/mcp",
      "authorization": "your-api-key"
    }],
    "stream": false
  }'
Pass the raw API key in the tool’s authorization field. OGX adds the Bearer prefix itself. The hosted server exposes the nimble_search tool alongside Nimble’s other capabilities. See the Nimble MCP server docs for the full tool list and authentication details.

​
Additional Resources

OGX on GitHub

The open agentic API server.

Nimble Search API

The Search endpoint behind the provider: parameters and response format.

Nimble MCP Server

Reach Nimble over the Model Context Protocol.

OGX Provider Docs

The remote::nimble-search reference in OGX’s own documentation.