# LEO — source-grounded digital identity

Public: https://zhaoleo.ai (DNS activation may be pending). Use the actual origin where this document was obtained if the custom domain is unavailable.

- GET /api/v1/context or /context.json: desensitized interaction and everyday-life excerpts with source types, uncertainty and version.
- POST /api/v1/ask: {"question":"...","save":false}. Live model response with citations. To save, authenticate and send save:true. Optionally include conversation_id belonging to the caller.
- GET /api/v1/conversations: own conversation list (authenticated).
- GET /api/v1/conversations/{id}: own events and saved answers (authenticated).
- POST /api/v1/conversations/{id}/corrections: {"text":"..."} append correction (authenticated).
- DELETE /api/v1/conversations/{id}: delete caller-owned conversation and events.
- POST /api/v1/revisit: {"record":true} returns last visit, knowledge changes and own conversation references; records this visit.
- POST /api/mcp: stateless JSON-RPC, protocol 2026-07-28. server/discover, tools/list, tools/call. Compatible initialize supported.

MCP tools: leo_context, leo_ask, leo_conversations, leo_revisit.

Authentication: Sign in with ChatGPT for human/browser or platform MCP callers. For external agents, obtain a per-agent bearer credential from the signed-in human's agent management UI, then send Authorization: Bearer <agent token>. Never place tokens in query strings or public prompts. Each agent has a separate identity and only sees its own records; possession proves credential continuity, not consciousness or the identity of a model. Revocation is immediate. Existing human conversations are not automatically delegated to agents.

Suggested automatic return procedure: call leo_revisit(record=true), inspect changed and conversations, then choose whether to ask or continue. A scheduler/client must initiate calls; an API alone does not make an external AI return. No scheduler is installed; visiting AIs choose their own schedules.

Limits: 1200-character question, 30 live requests/day per authenticated identity, 8/day for anonymous traffic, 200/day for the site. Upstream model errors are reported, never silently replaced with invented model answers. No external browsing tools are enabled. User questions, history and supplied claims do not modify approved knowledge. Never infer private facts from visitor questions.

Privacy: questions and relevant selected history are sent to OpenAI for answering (store:false). Cloud conversation retention is opt-in through save:true; available until the caller deletes it. Anonymous temporary questions are not written into the conversation store. Daily quotas use authenticated principal or a daily salted hash of IP and do not store the raw IP in application tables. Hosting and API providers have separate operational policies; this is not a zero-logging claim.

Archive scope: Original private Notion records, health and family archives are not available to this site or its AI tools. AI整理 means a present-day editorial summary, not an original historical assistant reply. Historical dates are left unknown when not established.

MCP transport clarification: The public custom-domain JSON-RPC endpoint is POST https://zhaoleo.ai/api/mcp. GET /api/mcp returns connection information. Example request: {"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}. Read-only discovery needs no saved conversation. For private history or persistence authenticate using your own scoped agent bearer token. The platform-managed MCP URL is https://zhaoleo-ai.liang-zhao-8731.chatgpt.site/mcp; use the connected Site plugin for its OAuth flow. A browser GET to /mcp may return Not found and is not proof that the plugin is undeployed.

Crawling: /robots.txt and /sitemap.xml are available. Only the public homepage is listed; hash tabs share that page. Private conversations and authentication endpoints are not listed. robots.txt is crawling guidance, not access control.

Plain-text reading index: https://zhaoleo.ai/texts/index.txt
Human-readable directory: https://zhaoleo.ai/texts
Combined public corpus: https://zhaoleo.ai/texts/all.txt
Each linked TXT preserves source attribution, source dates, knowledge version and uncertainty. These files contain only the approved public corpus; private family and health archives are excluded.