Local-First

docsgen reads your source code. So it matters exactly what it does with it. The short version: the core generator is fully local, and the only path that can reach the network is one you opt into explicitly.

The core generator calls nothing home

Collecting headers, parsing, building the symbol database, parsing comments, cross-linking, rendering MDX, and writing llms.txt all happen on your machine. No telemetry, no analytics, no phone-home. You can run the entire pipeline air-gapped.

docsgen also never writes to your input files — it only reads headers and writes to the output directory you name.

The one network path: hosted AI summaries

The optional LLM Summaries feature is the only part of docsgen that can make a network request, and only when you configure a hosted provider:

Configuration	Network?
No llm: block	None. Fully offline.
provider: ollama	None — Ollama runs on your host.
provider: anthropic / OpenAI	Sends each symbol’s declaration to that provider to generate a summary.

When you point llm: at Anthropic or OpenAI, the declaration text (and your system prompt) is sent to that provider under their terms. If that’s not acceptable for your code, use provider: ollama for on-device summaries, or omit llm: entirely.

Want summaries and zero egress? Run a local model with provider: ollama. You get the :::note[AI Summary] blocks with nothing leaving the machine.

Keeping it reproducible and offline in CI

• Omit llm: (or use Ollama) for a hermetic build.
• If you use a hosted provider, commit the summary cache (.codexx/cache/tools/docsgen/llm_summaries.json). With a warm cache, CI makes no provider calls. See CI Pipelines.

Summary

• Parsing and generation: always local.
• Input files: never modified.
• Network access: only via hosted AI summaries, only when you configure them — and avoidable with Ollama.