Agent discovery

LLMBase publishes machine-readable discovery metadata on https://llmbase.ai so AI agents can find the public website APIs, inference APIs, and MCP service without scraping HTML.

Discovery entrypoints

• /.well-known/api-catalog — org-wide API catalog in application/linkset+json
• /.well-known/openid-configuration — root OIDC discovery alias
• /.well-known/oauth-authorization-server — OAuth 2.0 authorization server metadata
• /.well-known/oauth-protected-resource — protected resource metadata for same-origin website APIs
• /.well-known/mcp/server-card.json — LLMBase MCP server card
• /.well-known/agent-skills/index.json — first-party Agent Skills discovery index

API descriptions

• /.well-known/service-desc/website-public.openapi.json
• /.well-known/service-desc/inference.openapi.json

These service descriptions are linked from the API catalog and the homepage’s HTTP Link headers.

Status endpoints

• /.well-known/status/website-api.json
• /.well-known/status/inference.json
• /.well-known/status/mcp.json

Agents can use these lightweight JSON endpoints instead of scraping marketing pages for health information.

Agent Skills

LLMBase hosts first-party skill artifacts directly on the website so agents can verify them with the SHA-256 digests from the discovery index.

Current skills:

• llmbase-api
• llmbase-model-discovery
• llmbase-mcp

WebMCP

On public pages, LLMBase registers a small set of read-only WebMCP tools for:

• public model search
• public leaderboard queries
• docs lookup
• pricing page lookup

These tools are only registered when the browser exposes a supported WebMCP API surface. Unsupported browsers continue to work normally.

Markdown for agents

LLMBase is prepared for Cloudflare’s Markdown for Agents negotiation. Public HTML responses vary on Accept, which is required for Cloudflare’s markdown conversion feature.

To serve text/markdown responses in production, the Cloudflare zone must have Markdown for Agents enabled for llmbase.ai.