# Docs

- [Overview](/docs): Two ways to run models with CarbonForge.
- API endpoint
  - [Quick start](/docs/endpoint): Send your first request to the CarbonForge endpoint.
  - [Reasoning](/docs/endpoint/reasoning): Controlling thinking depth, output volume, and latency.
  - [Streaming](/docs/endpoint/streaming): Token streaming, and the reasoning-first caveat for UIs.
  - [Errors](/docs/endpoint/errors): Every error shape the endpoint returns, and the failures that are not errors.
- Self-hosted container
  - [Quick start](/docs/container/quickstart): Pull the cf-bundle container and serve a tuned recipe on your own GPU.
  - [Context keys](/docs/container/context-keys): How a context key selects a benchmarked serving recipe.
  - [Configuration](/docs/container/configuration): Volumes, environment variables and launch modes for cf-bundle.
  - [Data boundary](/docs/container/data-boundary): Exactly what reaches CarbonForge, and what never leaves your infrastructure.
  - [Tracing](/docs/container/tracing): Per-request traces to a file or an OpenTelemetry collector.
  - [Troubleshooting](/docs/container/troubleshooting): Common failures, what they look like, and the fix.
- Tracing
  - [What tracing is](/docs/tracing): One record per request, at a privacy level you choose.
  - [Standard level](/docs/tracing/standard): Timings, token counts and scheduler metadata. No message content.
  - [Full level](/docs/tracing/full): The standard record plus raw prompt and output text.
- [For agents](/docs/agents): Machine-readable routes for coding agents and LLMs.