Originally created by: kumaakh
Suggested in PR [#133]. Fleet supports Claude, Gemini, Codex, and Copilot — the project context layer should be equally discoverable by all of them, not just Claude Code.
The llms.txt spec (llmstxt.org) defines a standard markdown file at the repo root that AI agents use to understand a project. Format:
```markdown
Brief summary blockquote
Body paragraphs explaining the project.
The full variant uses an XML-wrapped structure (same pattern as FastHTML's `llms-ctx-full.txt`, the reference implementation). Each document is wrapped in `<doc>` tags so the LLM knows where one ends and another begins:</doc>
```xml
<project title="Apra Fleet" summary="MCP server for orchestrating multiple agentic AI instances..."></project>
<docs>
<doc title="User Guide" desc="Installation, member registration, multi-provider setup">
[content of docs/user-guide.md]
</doc></docs>
<doc title="Vocabulary" desc="Fleet-specific terminology reference">
[content of docs/vocabulary.md]
</doc>
<doc title="Provider Matrix" desc="LLM provider capability comparison">
[content of docs/provider-matrix.md]
</doc>
<doc title="FAQ" desc="Common questions and answers">
[content of docs/FAQ.md]
</doc>
<doc title="Architecture" desc="Internal architecture and how fleet works">
[content of docs/architecture.md]
</doc>
```
Single source of truth: humans maintain `CLAUDE.md` and `llms.txt`.
`CLAUDE.md` is the Claude Code agent context file — it contains agent-specific instructions and is separate from `llms.txt`. They serve different purposes and should remain independent files.
Originally posted by: kumaakh
Fixed in commit 211f043 (PR [#128]).
llms.txtadded to repo root and maintained as the human-edited index.llms-full.txtis generated by CI on every build viascripts/gen-llms-full.mjsand committed to root — readable without a build step.Related
Tickets:
#128Ticket changed by: kumaakh