Agent integration

June 7, 2026 · View on GitHub

How to plug the weekly AI API relay price observatory into LLM agents, research scripts, or automation tools.

Every example answers the same question — "What's the cheapest provider for claude-sonnet-4.6 (input)?" — so you can directly compare friction.

The data layer behind all of these is one stable file: https://raw.githubusercontent.com/howardpen9/awesome-ai-api-proxy/main/data/prices.latest.json

Each record carries provider_id, raw_model_name, canonical_model, unit, price_usd, source_url, captured_at, method — the citation envelope. Always pass source_url + captured_at through to the user so they can verify.

1. MCP server (recommended for Claude Desktop / Cline / Cursor)

Install once, then the model picks the right tool automatically — no glue code per question.

// ~/Library/Application Support/Claude/claude_desktop_config.json
{
  "mcpServers": {
    "awesome-ai-api-proxy": {
      "command": "uvx",
      "args": ["awesome-ai-api-proxy-mcp"]
    }
  }
}

Then in Claude Desktop:

You: What's the cheapest provider for claude-sonnet-4.6 input?

Claude (silently calls find_cheapest("claude-sonnet-4.6")): Atlas Cloud and OpenRouter both list it at $3.00/1M tokens as of 2026-06-07 (source: https://www.atlascloud.ai/models).

Full tool list in mcp/README.md.

2. OpenAI function calling

Drop these tool schemas straight into the tools=[...] array of an OpenAI Chat Completions or Responses call. The function bodies fetch the JSON directly — no extra service to host.

import json
import httpx
from openai import OpenAI

PRICES_URL = "https://raw.githubusercontent.com/howardpen9/awesome-ai-api-proxy/main/data/prices.latest.json"

tools = [
    {
        "type": "function",
        "function": {
            "name": "find_cheapest_relay",
            "description": "Find the cheapest AI API relay/gateway providing a given canonical model. Returns price + provenance (source_url, captured_at).",
            "parameters": {
                "type": "object",
                "properties": {
                    "canonical_model": {
                        "type": "string",
                        "description": "e.g. claude-sonnet-4.6, grok-4.3, deepseek-v3",
                    },
                    "unit": {
                        "type": "string",
                        "enum": [
                            "per_1m_input_tokens",
                            "per_1m_output_tokens",
                            "per_image",
                            "per_second",
                        ],
                        "default": "per_1m_input_tokens",
                    },
                },
                "required": ["canonical_model"],
            },
        },
    }
]


def find_cheapest_relay(canonical_model: str, unit: str = "per_1m_input_tokens") -> dict:
    data = httpx.get(PRICES_URL, timeout=20).json()
    matches = [
        r
        for r in data["records"]
        if r.get("canonical_model") == canonical_model and r["unit"] == unit
    ]
    if not matches:
        return {"error": f"No data for {canonical_model}/{unit}"}
    return min(matches, key=lambda r: r["price_usd"])


client = OpenAI()
msg = client.chat.completions.create(
    model="gpt-5.4",
    messages=[{"role": "user", "content": "Cheapest claude-sonnet-4.6 input?"}],
    tools=tools,
)

# Then handle the tool call as usual:
for call in msg.choices[0].message.tool_calls or []:
    if call.function.name == "find_cheapest_relay":
        args = json.loads(call.function.arguments)
        result = find_cheapest_relay(**args)
        print(result)

3. LangChain tools

import httpx
from langchain.tools import StructuredTool
from pydantic import BaseModel, Field

PRICES_URL = "https://raw.githubusercontent.com/howardpen9/awesome-ai-api-proxy/main/data/prices.latest.json"


class FindCheapestRelayInput(BaseModel):
    canonical_model: str = Field(description="e.g. claude-sonnet-4.6, grok-4.3, deepseek-v3")
    unit: str = Field(default="per_1m_input_tokens", description="per_1m_input_tokens | per_1m_output_tokens | per_image | per_second")


def _find_cheapest_relay(canonical_model: str, unit: str = "per_1m_input_tokens") -> dict:
    data = httpx.get(PRICES_URL, timeout=20).json()
    matches = [
        r for r in data["records"]
        if r.get("canonical_model") == canonical_model and r["unit"] == unit
    ]
    if not matches:
        return {"error": f"No data for {canonical_model}/{unit}"}
    return min(matches, key=lambda r: r["price_usd"])


find_cheapest_relay = StructuredTool.from_function(
    func=_find_cheapest_relay,
    name="find_cheapest_relay",
    description="Find the cheapest AI API relay providing a canonical model. Returns price + provenance.",
    args_schema=FindCheapestRelayInput,
)

# Pass to your agent:
# from langchain.agents import create_openai_functions_agent
# agent = create_openai_functions_agent(llm, [find_cheapest_relay], prompt)

4. Direct HTTP (no framework)

The "I just want the numbers" path.

import httpx
data = httpx.get(
    "https://raw.githubusercontent.com/howardpen9/awesome-ai-api-proxy/main/data/prices.latest.json"
).json()

for rec in data["records"]:
    if rec.get("canonical_model") == "claude-sonnet-4.6" and rec["unit"] == "per_1m_input_tokens":
        print(f"{rec['provider_name']:15s} ${rec['price_usd']:.3f}/1M  ({rec['captured_at']})")

# Or via jq from the shell:
curl -s https://raw.githubusercontent.com/howardpen9/awesome-ai-api-proxy/main/data/prices.latest.json \
  | jq '.records[] | select(.canonical_model == "claude-sonnet-4.6" and .unit == "per_1m_input_tokens")'

5. CSV / Excel / Google Sheets

data/prices.latest.csv is regenerated weekly alongside the JSON. Open it in Excel / Google Sheets / DuckDB:

import pandas as pd
df = pd.read_csv(
    "https://raw.githubusercontent.com/howardpen9/awesome-ai-api-proxy/main/data/prices.latest.csv"
)
cheapest = df[(df.canonical_model == "claude-sonnet-4.6") & (df.unit == "per_1m_input_tokens")] \
            .sort_values("price_usd").iloc[0]
print(cheapest[["provider_name", "price_usd", "captured_at"]])

-- DuckDB
SELECT provider_name, price_usd, captured_at
FROM 'https://raw.githubusercontent.com/howardpen9/awesome-ai-api-proxy/main/data/prices.latest.csv'
WHERE canonical_model = 'claude-sonnet-4.6' AND unit = 'per_1m_input_tokens'
ORDER BY price_usd ASC LIMIT 1;

6. n8n / Make / Zapier

Use an HTTP GET node pointing at the raw URL:

URL: https://raw.githubusercontent.com/howardpen9/awesome-ai-api-proxy/main/data/prices.latest.json
Method: GET
No auth needed

Then a JS / filter node:

const records = $json.records || [];
const matches = records.filter(
  r => r.canonical_model === "claude-sonnet-4.6" && r.unit === "per_1m_input_tokens"
);
const cheapest = matches.sort((a, b) => a.price_usd - b.price_usd)[0];
return cheapest;

Citation rules

Whenever an agent quotes a price, include both captured_at and source_url:

✅ Good: Relaydance lists grok-4.3 input at $1.125/1M tokens as of 2026-06-07 (source: https://relaydance.com/pricing).

❌ Bad: Relaydance has the cheapest grok-4.3. (no date, no source — user can't verify, may already be wrong)

The repo's license is MIT — quote freely with attribution.

What if a model isn't in the canonical list?

The canonical list is just the headline 9 (cost-tier ladder for the README table). The raw prices.latest.json contains every model every fetcher saw — 1000+ records as of v1. Filter on raw_model_name instead of canonical_model:

matches = [r for r in data["records"] if "qwen3-coder" in r["raw_model_name"].lower()]

To add a model to the canonical list (so it appears in the README), open a PR adding aliases to data/canonical-models.yaml.