Build AI-powered Forge apps with Atlassian-hosted LLMs

LLMs have become a core tool in every app developer’s stack, powering features that feel almost magical. Users can query data with natural language, summarize long issue threads, and get smart, contextual suggestions, in an instant. Just a few years ago, these features would have been cumbersome or even impossible for developers to build.

AI features clearly provide a ton of useful functionality for app users, but developers integrating LLMs have to deal with trust considerations and complexity. Where is data going and how is it moving between systems? Is my model infrastructure scalable and compliant?

That’s why we’re excited to introduce the Forge LLMs API. The Forge LLMs API allows developers to call large language models hosted natively on Forge. Data stays on-platform, reducing risk and operational overhead. And because requests are routed only through Atlassian-managed infrastructure, apps that use Forge LLMs can still qualify for the Runs on Atlassian badge.

“With the Forge LLMs API, we’ve been able to give our customers something they couldn’t get anywhere else: an AI that orchestrates their entire delivery lifecycle inside Jira, with no data leaving their Atlassian tenant. That data protection assurance has opened doors with enterprise and government customers that were previously closed to AI tooling entirely.”
— Brett Taylor, CEO, BluBees

Try it out

It’s easy to get started. Because Forge handles model integration for you, you can skip the set-up and get straight to building AI-powered features. Just select which model you want to use, and start making calls using the LLMs SDK with just a few lines of code.

First, enable the model you want to use in your manifest.

modules:
  llm:
    - key: my-llm
      model:
        - claude

Install the SDK and import the client.

npm install @forge/llm

import { chat, stream } from '@forge/llm';

And you’re ready to send your first chat request.

try {
  const response = await chat({
    model: 'claude-sonnet-4-5-20250929',
    messages: [
      {
        role: 'user', content: 'Write a short poem about Forge LLMs.'
      }
    ],
  });

  console.log("#### LLM response:", JSON.stringify(response));
} catch (err) {
  console.error('#### LLM request failed:', { error:  err.context?.responseText });
  throw err;
}

Model availability

Under the hood, Forge LLMs are hosted on Bedrock, Amazon’s foundation layer for generative AI. Today, you can choose from three Claude models: Haiku, Sonnet, and Opus. Bedrock supports models from multiple providers, which means we’re exploring more models that could be brought online in the future.

You can choose which Claude model you want to use with each request, allowing you to balance costs, latency and capabilities, depending on your needs. Opus offers the best complex reasoning, but comes with higher costs and latency, while Haiku is fast and efficient for lightweight tasks. Sonnet sits in the middle ground between the two.

Inspiration and examples

You can find concrete, working examples and guided tutorials in the Forge LLMs docs to help spark ideas and get started quickly. Try it yourself by cloning one of these use cases:

• The Confluence AI assistant: Add AI-powered summarization and explanation to Confluence pages.
• LLM web trigger: Expose the Forge LLM service as an HTTP endpoint for external integrations.
• Agentic LLM web trigger: Give your LLM function-calling capabilities so it can interact with external tools and APIs.
• Long-running LLM with Realtime: Handle long-running LLM responses with async functions and real-time UI updates.

And that’s just the starting point. With the Forge LLMs API, you can build all kinds of transformative features for your users, from AI triage for Jira work items and pull requests, to agent-powered automations.

Pricing and observability

With the Preview release, LLMs are now a billable Forge capability. LLM usage is charged to the developer of the Forge app, and usage starting on June 1, 2026 will be counted toward your next Forge monthly bill. Pricing is usage based, so you only pay for what you use.

Forge LLM usage is tracked in credits, which correspond to model input and output tokens. Each model has a token-to-credit conversion ratio, and more powerful models use more credits per token. On your bill you’ll see two line items: input credits and output credits. You can also see a detailed breakdown of usage per model in the developer console.

Learn more about pricing

Ready to start building?

The Forge LLMs API now makes it easy to ship AI-powered features that meet enterprise expectations for trust, governance, and performance, without standing up your own model infrastructure or sending data off the Atlassian platform. You can spend less time on setup and more time building features that solve real-world problems.

We’re excited to see how you use the Forge LLMs API to deliver smarter, high-impact features for your users. As you ship real experiences powered by Forge LLMs, share what you’ve built with the developer community. We can’t wait to hear how Forge LLMs are showing up in your apps.