Claude API integration for products and internal tools

Anthropic's Claude API takes text, images and PDFs and returns text, structured JSON or calls to tools you define. We build it into web products and internal tools, along with the parts around the model: the tools it may call, schema checks, caching, cost and rate-limit handling, and a fallback when a request is declined.
Anthropic Claude
  • Tool use that connects Claude to your own data and actions
  • JSON output that matches your schema
  • Answers from your documents, with citations to the source passage
  • Prompt caching, batches and rate limits planned from the start

When Claude is the right choice

Claude suits work where the model reads a lot and must act on it carefully:

  • You need answers drawn from your own documents, with each claim pointing to the passage it came from. The citations feature does that for plain text, PDFs and your own content blocks.
  • Your application calls tools in a loop, such as a database lookup and then a report. Claude returns a structured tool call, your code runs it and sends back the result.
  • You reuse a long system prompt, tool list or document across many requests. Prompt caching lets later requests reuse that prefix.
  • You must keep traffic inside AWS or Google Cloud. Claude is offered through Amazon Bedrock and through Google Cloud, with your cloud's own credentials and billing.

When Claude is not the right choice

Plan around these before you commit:

  • You need to generate or edit images. Anthropic's vision page says Claude understands images but cannot create or edit them.
  • You need to name people in photos, exact object counts, or a reading of diagnostic scans. Anthropic lists these as vision limitations.
  • You need batch processing, the Files API or image and PDF URLs on Amazon Bedrock or on Claude models that Google manages. Anthropic lists these as not supported there; requests use base64 data instead.
  • You need citations and strict JSON in the same response. Anthropic's documentation says the two are incompatible and the API returns an error.

What we build with Claude

Typical projects, described in general terms.

Analysis written from your data

Claude reads records your system has already gathered and writes a plain-language diagnosis or summary, with the prompt open to the people who use it.

Assistants that call your tools

Tool definitions for lookups and actions in your own systems, run by your code, with checks before anything changes.

Document answers with citations

Questions answered from contracts, manuals or reports, with the cited passage shown next to each answer.

Structured extraction

Fields pulled from emails, forms, images and PDFs into JSON that matches a schema your code can rely on.

Batch jobs

Large overnight runs, such as classifying or summarizing many records, sent through the Message Batches API.

Guardrails and screening

Input screening, untrusted content kept in tool results, narrow permissions and handling for declined requests.

What the Claude API gives you

The Messages API takes a conversation and returns Claude's reply. With tool use, you describe functions with a JSON schema. When Claude wants one, the response stops with a tool_use block, your code runs it and sends a tool_result back. Some tools, such as web search and code execution, run on Anthropic's side instead. Structured outputs constrain the reply to your JSON schema, and strict tool definitions do the same for tool inputs. Claude reads JPEG, PNG, GIF and WebP images, and PDFs, where each page is read as text and as an image. Citations return the exact passage behind each claim. The Message Batches API processes large sets of requests asynchronously at a reduced cost. Prompt caching lets requests resume from a prefix that was already processed.

What a Claude integration needs beyond the API

  • Schema limits: structured outputs reject some JSON Schema features, such as numeric and string-length constraints, with an error. A refusal or a reply cut off at max_tokens may not match the schema, so we still check every result. The first request with a new schema is slower while it compiles.
  • Caching: the cache follows the order tools, system prompt, then messages. A change at one level clears that level and everything after it, and the default lifetime is five minutes, with a one-hour option. We order prompts so the stable parts come first, and read the cache fields in usage to confirm it works.
  • Rate limits: limits are counted in requests, input tokens and output tokens per minute, by usage tier. A 429 response can carry a retry-after header, and a sharp rise in traffic can hit acceleration limits, so we ramp traffic up gradually. For most current models, cached input does not count toward the input token limit.
  • Declined requests: current models include safety classifiers that can return a normal response with stop_reason: "refusal". We handle it in code, and where it suits, retry on another model.
  • Untrusted content: Anthropic recommends passing emails, web pages and documents in tool results, telling Claude where they came from and giving it only the access it needs. We test with content that contains injection attempts before launch.
  • Cloud choice: on Amazon Bedrock and Google Cloud, quotas and logging are run by that cloud, and some features are missing. We check the feature list for your platform before the design is fixed.

How we work

  1. Pick the platform first: the Claude API, Amazon Bedrock or Google Cloud, because features, quotas and billing differ
  2. Define the tools and schemas: what Claude may read, what it may change, and the shape of every result
  3. Test with real inputs: your own documents and records, including bad ones, refusals and injection attempts
  4. Hand over properly: prompts, caching layout, rate-limit handling and cost monitoring documented, with notes on reading a failed request

Frequently asked questions

Can Claude return data our code can parse reliably?

Yes. Structured outputs constrain the reply to your JSON schema, and strict tool use does the same for tool inputs. A refusal or a reply cut off by the token limit may not match, and some schema features are not supported, so results are still checked.

Can we use Claude through AWS or Google Cloud?

Yes. Claude is offered through Amazon Bedrock and through Google Cloud, using that cloud's credentials and billing. On Amazon Bedrock and on the Google-managed offering, some features, such as message batches, the Files API and image or PDF URLs, are not supported, and the cloud provider manages the quotas.

Can Claude answer from our own documents?

Yes. You can send plain text, PDFs or your own content blocks and turn on citations, so each claim points to the passage it came from. Scanned PDFs without extractable text cannot be cited, and citations cannot be combined with structured outputs.

How do we keep costs and rate limits under control?

Prompt caching reuses a long prompt prefix across requests, and for most current models cached input does not count toward the input token limit. Work that can wait goes through the Message Batches API at a reduced cost. Spend and rate limits can also be set per workspace.

How can we work together?

There are three ways, on the same terms:

  • Dedicated team, best for ongoing work: interview the developers, then they work on your project full or part time.
  • Fixed scope, best when requirements are clear: we agree the scope and the price up front, and deliver against it.
  • Time and material, best when the scope will change: you pay for the time used and can change priorities as the work goes on.

Whichever you choose, we sign an NDA before work starts, you own the code we write, and every change is reviewed in your repository.

Can I interview the developers before we start?

Yes. For a dedicated team you meet the developers in an online call and interview them before they join your project.

Building a feature on Claude?

Tell us what Claude should read, what it may do and which cloud you use. We will say what fits and what the integration involves.

Tell us about your project

Tell us what you need built, fixed or improved. We come back with questions, a suggested approach and an estimate.