Claude API Integration: How to Connect Anthropic's Claude (and OpenAI Models) to Your Company's Systems
A Claude API integration is a small backend service that your company runs between its own systems and the model. Employees sign in with the company login. The service holds the API credential, decides which data and tools the model gets, logs every call and enforces limits. The model never talks to your ERP directly, and no API key ever sits in a browser or a mobile app.
Three decisions shape everything else: the access route (the Anthropic API directly, or Claude through your cloud account), which data may leave your network, and what the model is allowed to do. An OpenAI API integration follows the same structure. Product details were checked against the official documentation, linked at the end, in November 2026.
Access routes: Anthropic API, Amazon Bedrock or Google Vertex AI
The route decides who operates the service, who processes your data and where the invoice comes from.
- Anthropic API, direct. You hold an account with Anthropic and issue API keys. Anthropic is the data processor, and you get the full feature set of the Claude API.
- Amazon Bedrock. AWS operates the service and is the data processor. Access runs through AWS IAM, usage is on your AWS invoice, and AWS states that model providers cannot see your prompts and completions.
- Google Cloud Vertex AI. Google operates Claude as a partner model, with Google Cloud IAM and billing. Google now runs Vertex AI under the name Gemini Enterprise Agent Platform.
Claude is also available in Microsoft Foundry. The cloud platforms follow their own release schedule: depending on platform and endpoint, features such as the MCP connector, the Message Batches API or the Files API are not offered there. Check the feature page before you design around one.
For OpenAI models the choice is similar: the OpenAI API directly, or Azure OpenAI in Microsoft Foundry, where Microsoft hosts the models and states that prompts and completions are not available to OpenAI. Rule of thumb: pick the route where your identity management, contracts and billing already live, unless a feature or region requirement says otherwise. For the model question itself, see which AI is better for what.
How a Claude API integration is built
- A backend service between your systems and the model. It authenticates the user, builds the prompt, calls the model, runs tools and writes the log.
- Tool use (function calling). You describe a function with a name, a description and an input schema. The model executes nothing: it asks for a call, your code decides whether to run it and returns the result. OpenAI calls this function calling.
- Retrieval over company documents. A search index finds the passages relevant to a question and passes them to the model as context. Filter by the user's permissions at search time.
- MCP connectors. The Model Context Protocol is an open standard for connecting models to tools and data. You write a connector once, for your ERP or ticket system, and reuse it. The Claude API can call remote MCP servers through its MCP connector, and OpenAI's API supports remote MCP servers too.
A tool definition for the Claude API is short:
{
"name": "get_order_status",
"description": "Look up the status of a customer order by order number.",
"input_schema": {
"type": "object",
"properties": {
"order_id": { "type": "string" }
},
"required": ["order_id"]
}
}
For the business side, see how to integrate AI into your company and our guide to ERP and AI integration.
Identity and secrets
- API keys live in a vault. The backend reads the key from a secrets manager at runtime. No keys in code, images or chat messages. Rotate them, and use separate keys per application and environment.
- Prefer cloud identities. On Bedrock, Google Cloud and Azure the service can authenticate with a cloud identity (IAM role, service account, Microsoft Entra ID) instead of a long-lived key.
- SSO in front. Users sign in through your identity provider, and the service passes that identity on to tools.
- Least privilege for tools. A tool acts with the rights of the requesting user, not with a service account that sees everything. Start read-only, and require confirmation for anything that sends, pays or deletes.
In the Binance hack of October 2026, an internal AI tool with too much reach was one link in the attack chain.
Data boundaries: what may leave your network
Decide per data category what the model may see, and enforce it in the backend: masking, field filters, blocked document classes.
- Training. Anthropic states that API data is not used for model training without your express permission. OpenAI does not use API data for training by default.
- Retention. Anthropic offers a zero data retention arrangement on request. It does not cover every feature (batch processing and the Files API store data by design) or every model. OpenAI has a similar option for approved customers.
- Region. At the time of writing, the direct Claude API lets you pin inference to the US or leave it global. EU processing is available through the cloud routes: EU inference profiles on Amazon Bedrock, the EU multi-region endpoint on Google Cloud, EU data zone deployments on Azure. OpenAI also offers European data residency to eligible API customers.
Our article on GDPR and AI integration covers the legal side. If a data class may not leave your network at all, a local model on your own servers is the alternative.
Cost control in an Anthropic API integration
- Usage limits. The Claude Console supports spend limits and rate limits per workspace. Add your own quotas per user or team in the backend.
- Prompt caching. Long, stable parts of a prompt, such as instructions, tool definitions or a reference document, can be cached and reused at lower cost.
- Batch processing. Jobs that can wait, such as nightly classification, run asynchronously at a reduced price.
- A smaller model for simple tasks. Routing, extraction and short summaries rarely need the largest model.
- Less context. Send the retrieved passages, not the whole document, and cap the output length.
Monitoring and evaluation
Log every request: user, application, model, token counts, tools called, latency and errors. Decide separately whether prompt and answer text is stored, because it may contain personal data. The platforms help: Anthropic provides a usage and cost API and OpenTelemetry tracing, Bedrock has model invocation logging, Google Cloud has request-response logging.
Keep a fixed set of real questions with expected answers and run it whenever you change a prompt, a tool or the model. Models are retired on a published schedule, so a model change will come.
Common mistakes
- Keys in client code. A key in a web page, mobile app or public repository will be found and used by others.
- Agents with broad permissions. One service account with access to everything turns a prompt injection into a data leak.
- No logging. Without logs you cannot explain an answer, a bill or an incident.
- No rate limits. A loop in one script can use up the monthly budget overnight.
- Trusting retrieved text. Content from documents, emails and web pages is data, not instructions.
- Consumer chat tools as a workaround. See shared AI chats are public.
Checklist before go-live
- Access route chosen, data processing terms signed.
- All model calls go through one backend service.
- Keys in a vault, rotation tested, none in client code.
- SSO in front, user identity passed to tools.
- Every tool has an owner, a scope and a confirmation rule for write actions.
- Data categories, region and retention settings documented.
- Spend limit, rate limits and alerts configured.
- Logging on, evaluation set in place, model update plan written.
Frequently asked questions
What is a Claude API integration?
A connection between your own software and Anthropic's Claude models through the API. In a company it is usually a backend service that authenticates users, calls the model, gives it selected data and tools, and logs and limits usage.
Should we use the Anthropic API directly or Amazon Bedrock or Google Vertex AI?
Use the direct Anthropic API if you need the full feature set of the Claude API. Use Amazon Bedrock or Google Cloud if your identity management, billing and contracts already sit there or you need a regional endpoint.
Is our data used to train the model in an Anthropic API integration?
According to Anthropic's documentation, API data is not used for model training without your express permission. A zero data retention arrangement is available on request, with some features and models excluded.
How does an OpenAI API integration differ from a Claude API integration?
The architecture is the same: a backend service, tools, retrieval, limits and logs. The access routes differ (OpenAI API or Azure OpenAI), and tool use is called function calling. A well-built backend can use both providers.
How do we keep the cost of a Claude API integration under control?
Set spend and rate limits per workspace, cache the stable part of long prompts, run non-urgent jobs as batches and use a smaller model for simple tasks.
Bottom line
The model call is the easy part of a Claude API integration. The work is in what surrounds it: who may ask, what the model may see and do, where data is processed and what it costs. Settle the access route and the data boundaries first, then build one backend that every application uses.
Planning to integrate Claude into your company? As an AI Forwarder we integrate hosted models such as Anthropic Claude and OpenAI GPT, over the direct API or through Amazon Bedrock, Azure OpenAI and Vertex AI, with SSO, least-privilege access, data boundaries and usage and cost limits. For Paysafe we delivered AI Forwarder projects with an Anthropic integration. Read more about Claude and OpenAI integration, our AI integration services and local models, or talk to us.
Sources and further reading
- Anthropic: Claude API features overview
- Anthropic: current Claude models
- Anthropic: tool use and MCP connector
- Anthropic: prompt caching and batch processing
- Anthropic: API and data retention and data residency
- Anthropic: workspaces and rate limits
- Anthropic: Claude in Amazon Bedrock and Claude models on Google Cloud
- AWS: data protection in Amazon Bedrock and cross-region inference
- Google Cloud: Anthropic Claude models
- Microsoft: data, privacy and security for Azure OpenAI models and deployment types
- OpenAI: function calling and data controls
- Model Context Protocol