The LLM gateway that learns to skip the LLM.

from openai import OpenAI

client = OpenAI(base_url="http://localhost:4000/v1", api_key="...")  # was: your LLM provider
r = client.chat.completions.create(model="auto", messages=[{"role": "user", "content": "who covers R-3?"}])
# r.model_extra["pankhllm"]["model"] == "decision:coverage_owner"   (2 ms, no LLM call)

# or the pankhllm client:  pip install pankhllm
from pankhllm import Client
a = Client("http://localhost:4000").ask("who covers R-3?", tags=["private"])
print(a.text, a.model)

import OpenAI from "openai";

const client = new OpenAI({ baseURL: "http://localhost:4000/v1", apiKey: "..." });
const r = await client.chat.completions.create({ model: "auto", messages: [{ role: "user", content: q }] });

// or the pankhllm client:  npm install @pankhllm/client
import { PankhClient } from "@pankhllm/client";
const a = await new PankhClient("http://localhost:4000").ask(q, { tags: ["private"] });

// Node 18+, plain JavaScript, no package needed
const r = await fetch("http://localhost:4000/v1/chat/completions", {
  method: "POST", headers: { "content-type": "application/json" },
  body: JSON.stringify({ model: "auto", messages: [{ role: "user", content: "who covers R-3?" }] })
});
const v = await r.json();
console.log(v.choices[0].message.content, v.pankhllm.model);   // "decision:coverage_owner"

// with the official openai package (CommonJS)
const OpenAI = require("openai");
const client = new OpenAI({ baseURL: "http://localhost:4000/v1", apiKey: "..." });

// or the pankhllm client:  npm install @pankhllm/client
import { PankhClient } from "@pankhllm/client";
const a = await new PankhClient("http://localhost:4000").ask("who covers R-3?");

// Semantic Kernel / Microsoft Agent Framework: change the endpoint, keep your plugins and loop
var kernel = Kernel.CreateBuilder()
    .AddOpenAIChatCompletion("auto", new Uri("http://localhost:4000/v1"), apiKey: "...")
    .Build();

// or the pankhllm client:  dotnet add package Pankhllm
var pk = new Pankhllm.PankhClient("http://localhost:4000");
var a = await pk.AskAsync("who covers R-3?", new AskOptions { Tags = ["private"] });
Console.WriteLine($"{a.Text} via {a.Model}");

import "github.com/singhpratech/pankhllm/sdks/go"

pk := pankhllm.New("http://localhost:4000")
a, err := pk.Ask(ctx, "who covers R-3?", pankhllm.Options{Tags: []string{"private"}})
if err != nil { log.Fatal(err) }
fmt.Println(a.Text, a.Model)   // a.Model == "decision:coverage_owner"

// any OpenAI Go client works too: set the base URL to http://localhost:4000/v1

import dev.pankhllm.PankhClient;

var pk = new PankhClient("http://localhost:4000");
var a  = pk.ask("who covers R-3?", new PankhClient.Options().tags(List.of("private")));
System.out.println(a.text() + " via " + a.model());

// Spring AI / LangChain4j / the official OpenAI Java SDK: base URL http://localhost:4000/v1

use pankhllm_client::{Client, Options};

let pk = Client::new("http://localhost:4000");
let a = pk.ask("who covers R-3?", &Options { tags: vec!["private".into()], ..Default::default() }).await?;
println!("{} via {}", a.text, a.model);

// the router itself:  cargo install pankhllm  &&  pankhllm serve --port 4000

require "pankhllm"

pk = Pankhllm::Client.new("http://localhost:4000")
a  = pk.ask("who covers R-3?", tags: ["private"])
puts "#{a.text} via #{a.model}"

# ruby-openai works too:  OpenAI::Client.new(uri_base: "http://localhost:4000/v1")

use Pankhllm\Client;

$pk = new Client('http://localhost:4000');
$a  = $pk->ask('who covers R-3?', ['tags' => ['private']]);
echo $a['text'], ' via ', $a['model'];

// openai-php works too: OpenAI::factory()->withBaseUri('http://localhost:4000/v1')

#include "pankhllm.h"   /* single .c/.h, libcurl */

char *body = pankh_simple_body("who covers R-3?", NULL, "[\"private\"]", 0);
pankh_response r = pankh_chat("http://localhost:4000", body, NULL);
printf("%d %s\n", r.status, r.body);
pankh_free(r.body); pankh_free(body);

#include "pankhllm.hpp"   // header-only, libcurl

pankhllm::Client pk("http://localhost:4000");
std::string json = pk.chat(R"({"model":"auto","messages":[{"role":"user","content":"who covers R-3?"}]})");
std::cout << json << std::endl;

curl -sS http://localhost:4000/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{"model":"auto","messages":[{"role":"user","content":"who covers R-3?"}]}'

# the response carries "pankhllm": {"model": "decision:coverage_owner", "attempts": [...]}

# train on your files, then serve the bundle
python3 notebooks/run_trainer.py --skills ./skills --prompts agent.yaml --logs app.jsonl --teacher ollama
./pankh-work/bundle/run.sh        # http://localhost:4000/v1

# or just the router, from source
cargo install pankhllm && pankhllm serve --port 4000