API Integration Tools: Connecting AI Agents to REST and GraphQL APIs

The API Tool Pattern

Most agent workflows eventually need to call external APIs — fetching data from a CRM, creating tickets in a project tracker, sending notifications, or querying third-party services. Rather than building a separate tool for every API endpoint, you can build a generic API caller that the agent configures per request.

This post shows two approaches: a constrained approach with predefined API configurations, and a flexible approach with a generic HTTP tool.

Approach 1: Predefined API Configurations

The safest pattern pre-registers each API the agent can call, including auth credentials and allowed endpoints:

from dataclasses import dataclass
from typing import Optional
import httpx

@dataclass
class APIConfig:
    name: str
    base_url: str
    auth_header: str
    auth_value: str
    allowed_methods: set[str]
    allowed_paths: list[str]  # regex patterns
    timeout: int = 15

class APIToolkit:
    def __init__(self):
        self.apis: dict[str, APIConfig] = {}
        self.client = httpx.AsyncClient(timeout=15)

    def register_api(self, config: APIConfig):
        self.apis[config.name] = config

    async def call_api(
        self,
        api_name: str,
        method: str,
        path: str,
        params: dict = None,
        body: dict = None,
    ) -> str:
        import json
        import re

        if api_name not in self.apis:
            return f"Error: Unknown API '{api_name}'. Available: {list(self.apis.keys())}"

        config = self.apis[api_name]

        if method.upper() not in config.allowed_methods:
            return f"Error: Method {method} not allowed for {api_name}"

        path_allowed = any(
            re.match(pattern, path) for pattern in config.allowed_paths
        )
        if not path_allowed:
            return f"Error: Path {path} not allowed for {api_name}"

        url = f"{config.base_url.rstrip('/')}/{path.lstrip('/')}"

        try:
            response = await self.client.request(
                method=method.upper(),
                url=url,
                params=params,
                json=body,
                headers={config.auth_header: config.auth_value},
            )
            response.raise_for_status()

            data = response.json()
            result = json.dumps(data, indent=2, default=str)

            if len(result) > 5000:
                result = result[:5000] + "\n[Response truncated]"

            return result
        except httpx.HTTPStatusError as e:
            return f"API error: HTTP {e.response.status_code} - {e.response.text[:500]}"
        except Exception as e:
            return f"Request failed: {str(e)}"

Register APIs at startup with their credentials kept server-side:

import os

toolkit = APIToolkit()

toolkit.register_api(APIConfig(
    name="github",
    base_url="https://api.github.com",
    auth_header="Authorization",
    auth_value=f"Bearer {os.environ['GITHUB_TOKEN']}",
    allowed_methods={"GET"},
    allowed_paths=[
        r"/repos/[\w-]+/[\w-]+$",
        r"/repos/[\w-]+/[\w-]+/issues.*",
        r"/repos/[\w-]+/[\w-]+/pulls.*",
    ],
))

toolkit.register_api(APIConfig(
    name="slack",
    base_url="https://slack.com/api",
    auth_header="Authorization",
    auth_value=f"Bearer {os.environ['SLACK_TOKEN']}",
    allowed_methods={"GET", "POST"},
    allowed_paths=[
        r"/chat.postMessage$",
        r"/channels.list$",
        r"/conversations.history$",
    ],
))

The LLM never sees API keys. It only knows the API name and the paths it can call.

The Tool Schema for Predefined APIs

api_tool_schema = {
    "type": "function",
    "function": {
        "name": "call_api",
        "description": "Call a registered external API. Available APIs: github (GET repos, issues, PRs), slack (GET/POST messages, channels). Auth is handled automatically.",
        "parameters": {
            "type": "object",
            "properties": {
                "api_name": {
                    "type": "string",
                    "enum": ["github", "slack"],
                    "description": "Which API to call"
                },
                "method": {
                    "type": "string",
                    "enum": ["GET", "POST", "PUT", "PATCH", "DELETE"],
                    "description": "HTTP method"
                },
                "path": {
                    "type": "string",
                    "description": "API endpoint path, e.g. /repos/owner/repo/issues"
                },
                "params": {
                    "type": "object",
                    "description": "URL query parameters as key-value pairs"
                },
                "body": {
                    "type": "object",
                    "description": "Request body for POST/PUT/PATCH requests"
                }
            },
            "required": ["api_name", "method", "path"]
        }
    }
}

Approach 2: GraphQL Tool

For GraphQL APIs, the tool accepts a query string and variables:

class GraphQLTool:
    def __init__(self, endpoint: str, headers: dict):
        self.endpoint = endpoint
        self.headers = headers
        self.client = httpx.AsyncClient(timeout=15)

    async def execute(self, query: str, variables: dict = None) -> str:
        import json

        if any(keyword in query.upper() for keyword in ["MUTATION", "DELETE"]):
            return "Error: Only queries are allowed, not mutations"

        try:
            response = await self.client.post(
                self.endpoint,
                json={"query": query, "variables": variables or {}},
                headers=self.headers,
            )
            data = response.json()

            if "errors" in data:
                return f"GraphQL errors: {json.dumps(data['errors'], indent=2)}"

            return json.dumps(data.get("data", {}), indent=2, default=str)
        except Exception as e:
            return f"GraphQL request failed: {str(e)}"

Retry Logic with Exponential Backoff

API calls fail. Build retry logic into your toolkit:

import asyncio

async def call_with_retry(
    func,
    max_retries: int = 3,
    base_delay: float = 1.0,
    **kwargs,
) -> str:
    for attempt in range(max_retries):
        result = await func(**kwargs)

        if not result.startswith("Error:") and not result.startswith("API error:"):
            return result

        if "HTTP 429" in result or "HTTP 5" in result:
            delay = base_delay * (2 ** attempt)
            await asyncio.sleep(delay)
            continue

        # Non-retryable error
        return result

    return result  # Return last error after all retries exhausted

Only retry on transient errors (429 rate limits, 5xx server errors). Client errors (400, 401, 404) should not be retried.

Response Mapping

API responses often contain more data than the LLM needs. Map responses to concise formats:

def map_github_issue(raw: dict) -> dict:
    return {
        "number": raw["number"],
        "title": raw["title"],
        "state": raw["state"],
        "author": raw["user"]["login"],
        "labels": [l["name"] for l in raw.get("labels", [])],
        "created": raw["created_at"],
        "comments": raw["comments"],
    }

Smaller, cleaner tool results mean less token usage and better LLM comprehension.

FAQ

Should the LLM know about API authentication details?

Never. API keys, tokens, and credentials must be stored server-side and injected by your tool implementation. The LLM should only know the API name and what endpoints are available. Exposing credentials risks leaking them through the LLM's output.

How do I handle paginated API responses?

Return the first page of results along with pagination metadata (next page token, total count). Let the LLM decide whether to fetch more pages by calling the tool again with pagination parameters. Do not automatically fetch all pages — this can result in hundreds of API calls.

When should I use predefined APIs vs a generic HTTP tool?

Use predefined APIs for production systems. They enforce strict access control and keep credentials safe. Use a generic HTTP tool only for development, prototyping, or internal tools where the user is trusted. In any system exposed to end users, always use the predefined approach.

---

#APIIntegration #REST #GraphQL #ToolDesign #AIAgents #AgenticAI #LearnAI #AIEngineering