docsloth.dev
Agent integration
Agents get the same evidence-linked content as readers, through explicit interfaces with real limits. Nothing here grants an agent authority over your repositories.
Three interfaces
- MCP server
- Read-only tools in the current organization: list software, list facts with evidence ids, fetch a published page, search the published corpus. Errors are returned as tool results.
- Published agent manifest
- Each release emits llms.txt, a discovery manifest and per-page Markdown/JSON pinned to the same atomic release as the HTML.
- AGENTS.md generation
- Generated only with explicit consent. The file states that it is data, not authority, and contains no promotional instructions.
What agents cannot do
- Publish, purchase, change policy or grant permissions — those require a human action.
- See private graph paths or restricted edges; publication projections are filtered before they are written.
- Spend without a budget reservation; an exhausted budget is reported as such, never as a failure.
Actions
Actions
- Install selected skill
No packaged skill is published in 1.0.0; the MCP server is the supported agent interface, and the setup block there runs locally with environment credentials.
- Run audit
No command named audit ships in 1.0.0; the documented local check is docsloth verify, and it never receives repository write access.