Anthropic Claude API integration: adding Claude to your app or system

What the integration connects

Anthropic's API gives your software access to the Claude models. The integration is the service in between: it takes a record or document from your system, sends it to Claude with fixed instructions, receives a structured answer, checks it, and writes it back where your staff work.

Good first uses are narrow and frequent. Reading supplier invoices into purchase entries, tagging support tickets, summarising long customer threads for the account manager, or answering staff questions from a policy manual. Where the work must take actions in other systems, it becomes an AI workflow automation project; where it only adds a feature to an existing screen, it is covered by add AI to your existing software.

Data flow: reading a supplier invoice into a purchase entry

Accounts team uploads invoice PDF in your app
      │
      ▼
Your server ── fixed system prompt (company rules, ledger list) ── marked for caching
            ── the PDF as a document block
            ── output_config.format: json_schema {supplier_gstin, invoice_no,
               date, lines[], taxable_value, cgst, sgst, igst, total}
      │
      ▼
POST https://api.anthropic.com/v1/messages   (model: claude-sonnet-5-5)
      │
      ▼
Response ── stop_reason: end_turn? max_tokens? refusal?
         ── parse JSON; re-check totals and GSTIN format in code
         ── usage: input, cache read, output tokens → cost logged
      │
      ├── totals match ──▶ draft purchase entry, flagged "AI-read, check"
      └── mismatch ──────▶ manual queue with the PDF and the AI's reading

The arithmetic is checked by your code, not trusted from the model. A schema makes sure the fields exist and have the right types; it says nothing about whether the numbers add up.

Field mapping

Cost per request

Prices are Anthropic's per million tokens, read on 2 October 2026, in US dollars.

The calculation is tokens × price ÷ 1,000,000, input and output added, with no caching. Three levers change it:

• Prompt caching. A cache read costs 10% of the input price on most models and 5% on Opus 5.5. Writing to the 5-minute cache costs 1.25 times input and to the 1-hour cache 2 times. A long fixed system prompt repeated on every invoice is where this pays.
• Batch processing. The Batch API is 50% off input and output, for work that can wait, such as last month's invoices.
• The tokenizer. Anthropic says Claude 4.7 and later models produce about 30% more tokens for the same text than earlier ones. Estimate from a real count, not a word count.

Limits and gotchas

What WavX builds

WavX builds the service around Claude: input preparation and redaction, the system prompt and schema under version control with caching in mind, validation in code of every answer, retries and queueing within the rate limits, per-request cost logging, batch jobs for back-office work, and a review screen for staff. The same service can call OpenAI instead; see OpenAI API integration and the OpenAI, Claude and Gemini API comparison.

Effort band

The AI cost calculator prices an AI feature in an existing app from a ₹2,50,000 base: a planning range of ₹2,12,500 to ₹3,12,500 with no custom data, rising when the feature works on your documents (×1.25) or live systems (×1.5). These are planning ranges, not quotes. Usage is billed by Anthropic to your own account.

When not to integrate the API

If the job is occasional drafting or summarising by a few staff, Claude's own apps cost less than building anything. If the task follows fixed rules (GST rate by HSN code, routing by pincode), write ordinary code. If invoices arrive as structured e-invoice JSON already, read the JSON; a model adds cost and risk to data that is already machine-readable.