# הפודקאסט של מיכאל — API

> Read-only HTTP + MCP surface for הפודקאסט של מיכאל. No write methods. Auth is optional (anonymous OAuth 2.1 + PKCE S256 available for clients that prefer a bearer token).

## Quickstart

Three lines from zero to a real episode:

```bash
curl https://podcast.lugassy.net/status                         # 1. health check
curl 'https://podcast.lugassy.net/api/search?q=ai&limit=3'      # 2. find an episode
curl https://podcast.lugassy.net/1.md                           # 3. read the transcript (replace 1)
```

## Authentication

**Auth is optional.** Every endpoint is public, read-only, and CORS-open. Two modes:

- **Zero-auth (default).** Just call the endpoints — no header, no signup.
- **Public OAuth 2.1 + PKCE S256.** For clients that prefer M2M / client_credentials flows or want a bearer token to log per-request quotas, anonymous tokens are issued without consent at `https://podcast.lugassy.net/oauth/token`. Discovery: `https://podcast.lugassy.net/.well-known/oauth-authorization-server` (RFC 8414) and `https://podcast.lugassy.net/.well-known/oauth-protected-resource` (RFC 9728).

### M2M / agent auth walkthrough (optional)

```bash
# 1. Generate a PKCE verifier + S256 challenge (any client lib does this)
# 2. Anonymous client_credentials grant — no client secret, public client
curl -X POST https://podcast.lugassy.net/oauth/token \
  -H 'Content-Type: application/x-www-form-urlencoded' \
  -d 'grant_type=client_credentials&client_id=public&scope=read:episodes search:episodes'
# Response: { access_token, token_type: "Bearer", expires_in, scope }

# 3. Use the token (or skip — anonymous calls also work)
curl -H 'Authorization: Bearer <token>' 'https://podcast.lugassy.net/api/search?q=ai'
```

Available scopes:

- `read:episodes` — read episode metadata, audio URLs, transcript URLs.
- `read:transcripts` — read full transcript text.
- `search:episodes` — call /api/search and /ask.

All scopes are granted automatically on anonymous client_credentials.

## SDK install

There's no proprietary SDK. Use any HTTP client; for MCP, any MCP-compliant client works:

```bash
# JavaScript / TypeScript
npm install undici          # or use built-in fetch

# Python
pip install httpx           # or requests

# MCP (Claude.ai, ChatGPT, Cursor)
# Add this URL as a custom MCP connector — no install:
# https://podcast.lugassy.net/mcp  (transport: streamable-http, auth: none)
```

## Rate limits

- **60 requests/minute per IP** across all API endpoints (`/api/*`, `/mcp`, `/.well-known/mcp`, `/ask`, `/status`).
- Every response carries `X-RateLimit-Limit`, `X-RateLimit-Remaining`, `X-RateLimit-Reset` (Unix seconds).
- 429 responses carry `Retry-After` (seconds). Self-throttle on those headers.

## Errors

Every error is a structured JSON envelope:
```json
{ "error": { "code": "episode_not_found", "message": "…", "hint": "…", "docs_url": "…" } }
```
Status codes: 400 (bad query/body), 402 (donation requested at /donate), 404 (no such episode), 405 (wrong method), 429 (rate-limited), 500 (server side).

## Endpoints

### Search
`GET https://podcast.lugassy.net/api/search?q=<query>&limit=<n>`

Ranked full-text search over episode title + description + transcript.
Response: `{ query, count, took_ms, results: [{ id, title, date, url, audio, transcript, score, snippet }] }`.

### Ask (NLWeb)
`POST https://podcast.lugassy.net/ask` — body: `{ "query": "...", "limit": 10 }`
`GET https://podcast.lugassy.net/ask?q=<query>&limit=<n>` — query-string variant

Natural-language ask. Returns episodes ranked by transcript relevance, wrapped in NLWeb `_meta` envelope.
Set `Accept: text/event-stream` (or `Prefer: streaming=true`) for SSE: events `start`, `result` (one per match), `complete`.

### Async (202 Accepted + polling)

Long-running operations can opt into the async pattern. Three interchangeable entry points: **`POST https://podcast.lugassy.net/jobs`** with body `{ kind: "ask"|"search", query, limit }` (conventional path), **`POST https://podcast.lugassy.net/ask?async=1`** (or `?async=1` on `/api/search`), or set **`Prefer: respond-async`** on any of the above. All return **HTTP 202 Accepted** with `Location: /jobs/<id>`, `Retry-After: 1`, and a JSON body `{ job_id, status: "pending", poll_url, retry_after_seconds }`. Poll `GET https://podcast.lugassy.net/jobs/<id>` until `status` flips from `"pending"` to `"completed"`; the final response carries the result under `.result`.

```bash
# 1. Kick off the job — 202 Accepted
curl -i -X POST 'https://podcast.lugassy.net/ask?async=1' \
  -H 'Content-Type: application/json' \
  -d '{"query":"ai agents"}'
# HTTP/1.1 202 Accepted
# Location: https://podcast.lugassy.net/jobs/<id>
# Retry-After: 1
# { "job_id": "<id>", "status": "pending", "poll_url": "https://podcast.lugassy.net/jobs/<id>", "retry_after_seconds": 1 }

# 2. Poll the job — 200 OK
curl https://podcast.lugassy.net/jobs/<id>
# { "job_id": "<id>", "status": "completed", "result": { "query": "...", "results": [...] } }
```

Job ids are stateless: the id encodes the spec, so any client that holds an id can resume polling later from a different IP / device.

### Status
`GET https://podcast.lugassy.net/status` — health snapshot for circuit-breaker logic.
Always 200 when reachable. Response includes show name, episode count, latest episode summary.

### MCP server (Streamable HTTP, JSON-RPC 2.0)
`POST https://podcast.lugassy.net/mcp` — tool calls
`GET https://podcast.lugassy.net/mcp` — manifest summary

Methods: `initialize`, `ping`, `tools/list`, `tools/call`, `resources/list`, `resources/read`.
Tools: `search_episodes`, `get_episode`, `get_latest_episode`.
Batch / bulk: the request body may be a JSON-RPC 2.0 batch — an array of up to 50 request objects run in one round-trip, answered by an array of responses in order.

MCP discovery URLs (all return the same manifest):
- https://podcast.lugassy.net/.well-known/mcp
- https://podcast.lugassy.net/.well-known/mcp.json
- https://podcast.lugassy.net/.well-known/mcp-configuration
- https://podcast.lugassy.net/.well-known/mcp/server.json

### OpenAPI
`GET https://podcast.lugassy.net/.well-known/openapi.json` — OpenAPI 3.1 spec for the entire read surface.

## Agent mode

Append `?mode=agent` to `/` or to any `/<id>` to get a compact JSON envelope with endpoint inventory and either the latest episode (homepage) or the specific episode (episode page).

## Markdown view

- `https://podcast.lugassy.net/index.md` — homepage as markdown
- `https://podcast.lugassy.net/<id>.md` — episode page as markdown
- Or send `Accept: text/markdown` on any HTML page.