Amazon Textract now plugs into Claude Code and Cursor as an agent skill
AWS has published a Textract Core skill that lets coding and chat agents extract text, tables and forms from documents, with no server to run and no fee beyond standard Textract pricing.

Amazon Web Services has added a skill for AI agents to Amazon Textract Core, its document-extraction service. According to the official announcement, an agent can now pull text, tables and forms out of documents without any server or infrastructure to deploy.
What the skill does for an agent
AWS describes the skill as a set of instructions that steers an agent toward the right Textract Core API for each document. It also bundles a parser that AWS calls tested. The parser turns raw Textract output into text in reading order, plus structured fields: answers to questions, receipt and invoice details, and identity document fields.

The user asks in natural language, and the agent returns the answers. Multi-page PDFs are supported, and every page comes back in reading order.
The skill also uses Textract’s confidence scores. Results with high confidence are accepted, and the rest are flagged so they can go to the customer’s own review process. AWS’s example is an agent that turns a patient intake form into structured data and marks unclear fields for staff to check.
Before this, AWS says, teams that wanted an agent to use Textract had to build and maintain their own integration. The company’s pitch is that the skill removes that work, with “no server to run or infrastructure to deploy.”
Who can use it, where, and at what price
The skill carries no extra charge. Standard Textract pricing applies to the documents an agent processes, so the cost sits in the API calls, not in the skill itself.
It is available in every commercial AWS Region that offers Textract. AWS lists 16:
- US: Ohio, N. Virginia, N. California and Oregon
- Asia Pacific: Mumbai, Seoul, Singapore, Sydney and Tokyo
- Canada (Central)
- Europe: Frankfurt, Ireland, London, Paris, Spain and Zurich
There are two ways to get it. Teams that use the AWS MCP server with the aws-core plugin can have their agent load the skill automatically. The skill is also published in the AWS Agent Toolkit on GitHub, so agents such as Kiro, Claude Code and Cursor can use it. AWS says “including,” so the list of supported agents is not stated as complete.
One more step in a document push
The analysis here is about track record. The skill is the second Textract announcement from AWS in two days that is aimed at agent-driven document work, and it follows a pattern of packaging existing services so agents can call them directly.
Our earlier report said Amazon Textract can now split a 3,000-page packet and score every value. That launch concerned how much a single job can digest and how confidence is reported. The new skill concerns who can call the service: any agent that supports skills, without custom glue code. Read together, they point at the same goal of making documents something an agent can handle end to end.
The same approach showed up in agent tooling elsewhere in AWS this week. Kimi Code CLI now signs in to Bedrock with your AWS login, no API key, which also removes a piece of integration work for developers. And on the pricing side, Amazon Connect will handle refunds and disputes for $0.40 a case. That shows AWS is willing to price agent work per outcome, while the Textract skill stays on the existing per-page style of billing and adds nothing on top.
The skill’s distribution is also notable. Listing Claude Code and Cursor next to AWS’s own Kiro suggests the company wants its services reachable from whichever agent a team already uses, not only from its own tools.
What the announcement leaves unproven
The page offers no accuracy figures, benchmark results or speed numbers, so the claim that the parser is tested rests on AWS’s word alone. It does not say which document types beyond PDFs are covered, or how the confidence threshold is chosen. It does not say whether teams can tune that threshold, which matters for the intake-form kind of workflow AWS uses as its example.
Review remains the customer’s job. The skill flags low-confidence fields, but routing them to a person, and deciding what counts as good enough for medical or identity data, is left to each team’s own process.
For developers, the practical change is modest but real: a document task that once needed a custom integration can be a one-line request to an agent that already has the skill. Whether the results hold up on messy real-world scans is something AWS has not shown, and teams will have to test it on their own documents. Full setup details are in the AWS Agent Toolkit repository on GitHub, which the announcement points to.
Sources
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Alex CarterAI News & Big Tech CorrespondentAlex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.
You may also like

Amazon Bedrock AgentCore adds a web search API at $7 per 1,000 queries

Amazon says PyTorch code now runs on Trainium chips with one setting changed

Codex can now run on Amazon Bedrock, with AWS logins instead of OpenAI keys
