> ## Documentation Index
> Fetch the complete documentation index at: https://docs.reasonblocks.com/llms.txt
> Use this file to discover all available pages before exploring further.

# What is ReasonBlocks?

> Connect your agent's model calls, collect workflow data and prepare a smaller model for your application.

ReasonBlocks captures the model calls in your agent's workflow and uses your
application's tasks to train and evaluate a smaller model. You keep your provider
credentials and your existing Python client library throughout.

OpenAI, Anthropic, Google Gemini and Amazon Bedrock applications all connect the same
way: a data source in the dashboard, and a small generated helper. Gemini and Bedrock have
[their own setup section](/client-integration#gemini-and-bedrock), because their SDKs
take the connection settings differently.

## Start with capture

For a Gemini or Bedrock application, read
[Gemini and Bedrock](/client-integration#gemini-and-bedrock) — you create the source
the same way, and that section shows the client setup.

Create a source in **Data**, copy its capture URL and key, and ask your coding
agent to read the [setup guide](/agent-setup) — its plain-text copy is
[/agent-setup.md](/agent-setup.md) — and integrate ReasonBlocks into the project.
The exact instruction to paste is at the top of that guide.

The [setup CLI and skill](/agent-setup) generate a small Python helper and project
configuration. Your coding agent connects the helper to the existing client and
labels each task's calls. The CLI and helper ship in `rbtrace==1.2.1` on PyPI;
the setup guide has the install commands.

Capture does not require a sandbox or start training. Normal provider charges
continue to apply when your application runs.

## Train when your application is ready

Full-agent training needs representative tasks and a test copy of your tools and
data that can reset between attempts. Your coding agent can help build the adapter
for that environment. ReasonBlocks then prepares a plan, trains within the approved
budget and reports complete-task results for review.

Enabling a trained release is a separate decision. See
[Train your complete agent](/full-agent-training) for the environment, evaluation
and serving requirements.

## Get started

<CardGroup cols={2}>
  <Card title="Set up with your coding agent" icon="terminal" href="/agent-setup">
    One instruction, a setup CLI and a reusable skill.
  </Card>

  <Card title="Quickstart" icon="rocket" href="/quickstart">
    Create a data source and connect your existing application.
  </Card>

  <Card title="Integrating your agent" icon="plug" href="/client-integration">
    Client settings, task headers and deployment checks.
  </Card>

  <Card title="Train your complete agent" icon="flask" href="/full-agent-training">
    Prepare a test environment, review training and evaluate a release.
  </Card>

  <Card title="Deploy your connected application" icon="server" href="/deployment">
    Ship the helper, connection configuration and runtime secrets.
  </Card>
</CardGroup>