Claude Skills API: What's New, How to Use It, and Alternatives (2026)
Three capabilities left beta on the same day, and one sentence in the announcement explains why they shipped together.
Anthropic writes that together they let you build agents that "operate software, apply your team's expertise, and return finished files."
Read that as three separate problems being closed at once. Seeing the screen, knowing your procedure, and handing back a document.
This guide covers what went generally available and what a skill actually is. It then covers what the release does not say, and where the same idea already reaches people who do not write code.
What went generally available on August 20
The announcement is dated August 20, 2026, and it covers three things.
Computer use, the Skills API, and the Files API are all generally available on the Claude Platform. Computer use also gained a new browser use tool.
The framing is worth noting. This is a platform release aimed at people building agents, not a new button inside a chat window.
If you are not building an agent, skip to the section on where this reaches you anyway. The concepts still matter, because they are showing up inside tools you already pay for.
What a skill actually is
The definition on the page is the clearest one Anthropic has published, and it is short.
A skill is "a folder of instructions, scripts, and templates that Claude loads only when a task calls for it."
Three details in that sentence do real work. It is a folder rather than a prompt, it can contain runnable scripts, and it is loaded on demand.
The last part is the one people miss. Nothing is loaded until a task calls for it, so a large library of skills does not tax every request.
The Skills API adds the plumbing around that folder. You "upload and version your own skills, then attach them to any request."
Hosting is not your problem. The page states that skills "run in Claude's code execution sandbox, so there is nothing for you to host."
The Box example in the announcement shows what this looks like in practice. A bank encodes its credit methodology and approved memo format as a skill, and the agent applies it to documents already in Box.
What the Files API changes
The Files API is storage for what an agent reads and writes.
The workflow described is simple. Upload a PDF or spreadsheet once, reference it by ID in later requests, then download whatever the agent produced.
That removes a specific waste. Without it, the same document gets re-sent with every request in a long conversation.
General availability brought three concrete changes: automatic file expiration, rate limits five times higher, and 1 TB of storage per organization.
Computer use and the browser use tool
Computer use lets an agent operate software it can see. Given a screenshot, it clicks, types, and scrolls the way a person would.
The stated point is coverage rather than elegance. It works in "applications that were never built for automation," which describes a great deal of enterprise software.
The new browser use tool narrows that for the web. Alongside the screenshot, the agent "reads the structure of the page and acts on a specific field or button rather than a position on screen."
That distinction matters for reliability. Targeting a named field survives a layout change, while targeting a pixel position does not.
What general availability actually improved
| Surface | What changed at GA |
|---|---|
| Computer use | Several actions per turn instead of one per model call |
| Computer use | Now eligible for HIPAA-regulated workloads under the BAA |
| Browser use tool | New today, with page structure plus multi-action turns |
| Skills API | Simpler uploading and versioning of your own skills |
| Files API | Automatic expiration, 5x higher rate limits, 1 TB per organization |
The multi-action turn is the change with the clearest payoff. Fewer model calls means fewer round trips, and the announcement's customer quote puts numbers on it.
One team reports a claims workflow dropping from 32 minutes to 13. The same quote cites cost per task falling about 30% and completion reaching 100%, with no prompt changes.
Treat that as one customer's result rather than a benchmark. It is specific, attributed, and not generalized by Anthropic into a claim about all workflows.
Where this reaches you without writing code
This is the part that matters if you are an analyst rather than an engineer.
The skill concept is not exclusive to one vendor's API. It is becoming the standard way tools package repeatable expertise.
Powerdrill Bloom is one place it already appears. Its plan pages list running built-in and custom agent skills, and using Claude Skills to run specialized workflows for research, analysis, and execution.
The practical version of that is unglamorous and useful. Your reporting rules stop living in a document nobody opens, and start living in something that runs.
Our guide to running Claude Skills outside Claude covers the portability question in more detail.
Alternatives, sorted by how much you want to build
Pick by the amount of engineering you are willing to own, not by which model you prefer.
| You want | Reasonable route |
|---|---|
| Full control of an agent you host | The Skills API and Files API directly |
| Packaged skills inside a product | Powerdrill Bloom, or a tool that supports Claude Skills |
| Skills in a different assistant | Grok Skills, which ships built-in document and spreadsheet skills |
| A vendor-neutral packaging format | The Agent Plugins specification |
| A standard way to connect tools and data | Model Context Protocol |
Two of those deserve a note. Grok ships its own built-in skills for documents, decks, and spreadsheets, so the idea is not unique to one platform.
The Agent Plugins specification tackles packaging rather than execution. Our explainers on agent plugins and what MCP is cover how those two standards divide the work.
If you want the assistant-level comparison instead, see the roundup of Claude Skills alternatives.
What the announcement does not say
Three gaps are worth naming, because filling them in with guesses is how stale articles get written.
There is no separate price for the Skills API or the Files API on the announcement page. Storage and rate limits are described, and no fee is attached to them.
There is no statement about skill size limits, script runtimes, or what the sandbox can reach. The documentation links exist for exactly that reason.
Platform coverage is stated but uneven. The Skills API and the Files API are also available through Microsoft Foundry.
The other two are not there yet. The updated computer use and browser use tools are described as coming soon to Google Cloud's Vertex AI.
One reassurance is explicit. Existing beta integrations keep working while you migrate.
That last line is the one to act on if you already built against the beta. There is no forced cutover date on the page, and no migration deadline is named.
Conclusion
The release closes a loop that was previously three separate chores. An agent can now see an application, apply your team's written procedure, and hand back a file.
For builders, the practical news is versioned skills you do not host and file storage you do not re-send. For everyone else, the useful takeaway is the shape of a skill.
Write your procedure down as instructions plus templates, and it stops being tribal knowledge. Would you rather use skills than build them? Then try Powerdrill Bloom and see the AI report generator page.
Frequently asked questions
What is the Claude Skills API?
It is an API for uploading and versioning your own skills, then attaching them to requests. A skill is a folder of instructions, scripts, and templates loaded only when a task needs it.
Do I need to host anything to run a skill?
No. Anthropic states that skills run in Claude's code execution sandbox, so there is nothing for you to host.
What does the Files API do?
It stores documents an agent reads and writes. You upload a file once, reference it by ID later, and download whatever the agent creates.
What is the browser use tool?
It is a new part of computer use for web applications. It reads page structure and acts on a named field or button rather than a screen position.
Can I use skills without building an agent?
Yes, through products that support them. Powerdrill Bloom's plan pages list built-in and custom agent skills alongside Claude Skills workflows.