@pinecall/skills 0.1.1 → 0.1.3

package/build.mjs

@@ -33,7 +33,9 @@ const HOUSE_RULES = `## House rules — always apply
 - **TTS model is auto-derived from \`language\`** — non-English agents (e.g.
   \`language: "es"\`) default ElevenLabs to \`eleven_multilingual_v2\` so numbers,
   dates and currency are pronounced correctly (flash/turbo don't normalize text).
-  English stays on \`eleven_flash_v2_5\`. Override with \`voice: { ..., model: "..." }\`.
+  English stays on \`eleven_flash_v2_5\`. To keep flash on a non-English agent
+  (lower latency/cost), set the top-level \`flash: true\` flag. To pin any model,
+  use \`voice: { ..., model: "..." }\` (explicit model always wins over \`flash\`).
 - **Greeting**: inbound → \`greeting\` field in \`pc.agent()\`; outbound → \`greeting\`
   field in \`agent.dial()\`. It is sugar for \`call.say()\` in \`call.started\`.
 - **Auth**: \`new Pinecall()\` reads \`PINECALL_API_KEY\` from env and auto-connects.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pinecall/skills",
-  "version": "0.1.1",
+  "version": "0.1.3",
   "description": "Agent Skills for the Pinecall SDK — installable into Claude Code, Antigravity, Cursor, Copilot and any agent that supports the open Skills format.",
   "type": "module",
   "license": "MIT",

package/skills/pinecall-agent-skills/SKILL.md

@@ -30,7 +30,9 @@ table below indexes every page; open the `references/…` file for the full text
 - **TTS model is auto-derived from `language`** — non-English agents (e.g.
   `language: "es"`) default ElevenLabs to `eleven_multilingual_v2` so numbers,
   dates and currency are pronounced correctly (flash/turbo don't normalize text).
-  English stays on `eleven_flash_v2_5`. Override with `voice: { ..., model: "..." }`.
+  English stays on `eleven_flash_v2_5`. To keep flash on a non-English agent
+  (lower latency/cost), set the top-level `flash: true` flag. To pin any model,
+  use `voice: { ..., model: "..." }` (explicit model always wins over `flash`).
 - **Greeting**: inbound → `greeting` field in `pc.agent()`; outbound → `greeting`
   field in `agent.dial()`. It is sugar for `call.say()` in `call.started`.
 - **Auth**: `new Pinecall()` reads `PINECALL_API_KEY` from env and auto-connects.

package/skills/pinecall-concepts/SKILL.md

@@ -34,7 +34,9 @@ table below indexes every page; open the `references/…` file for the full text
 - **TTS model is auto-derived from `language`** — non-English agents (e.g.
   `language: "es"`) default ElevenLabs to `eleven_multilingual_v2` so numbers,
   dates and currency are pronounced correctly (flash/turbo don't normalize text).
-  English stays on `eleven_flash_v2_5`. Override with `voice: { ..., model: "..." }`.
+  English stays on `eleven_flash_v2_5`. To keep flash on a non-English agent
+  (lower latency/cost), set the top-level `flash: true` flag. To pin any model,
+  use `voice: { ..., model: "..." }` (explicit model always wins over `flash`).
 - **Greeting**: inbound → `greeting` field in `pc.agent()`; outbound → `greeting`
   field in `agent.dial()`. It is sugar for `call.say()` in `call.started`.
 - **Auth**: `new Pinecall()` reads `PINECALL_API_KEY` from env and auto-connects.

package/skills/pinecall-examples/SKILL.md

@@ -52,7 +52,9 @@ const agent = pc.agent("mara", {
 - **TTS model is auto-derived from `language`** — non-English agents (e.g.
   `language: "es"`) default ElevenLabs to `eleven_multilingual_v2` so numbers,
   dates and currency are pronounced correctly (flash/turbo don't normalize text).
-  English stays on `eleven_flash_v2_5`. Override with `voice: { ..., model: "..." }`.
+  English stays on `eleven_flash_v2_5`. To keep flash on a non-English agent
+  (lower latency/cost), set the top-level `flash: true` flag. To pin any model,
+  use `voice: { ..., model: "..." }` (explicit model always wins over `flash`).
 - **Greeting**: inbound → `greeting` field in `pc.agent()`; outbound → `greeting`
   field in `agent.dial()`. It is sugar for `call.say()` in `call.started`.
 - **Auth**: `new Pinecall()` reads `PINECALL_API_KEY` from env and auto-connects.

package/skills/pinecall-guides/SKILL.md

@@ -26,6 +26,7 @@ table below indexes every page; open the `references/…` file for the full text
 | **Tools and Functions** | Let your agent take actions: look up data, transfer calls, book appointments. | [`references/guides/tools-and-functions.md`](references/guides/tools-and-functions.md) · [docs](https://docs.pinecall.io/guides/tools-and-functions) |
 | **Knowledge bases (RAG)** | Tutorial — ground a voice or chat agent on your own documents with retrieval-augmented generation. | [`references/guides/knowledge-bases.md`](references/guides/knowledge-bases.md) · [docs](https://docs.pinecall.io/guides/knowledge-bases) |
 | **Multi-Tenant Dashboards** | Host many tenants on one Pinecall instance with scoped event streams. | [`references/guides/multi-tenant.md`](references/guides/multi-tenant.md) · [docs](https://docs.pinecall.io/guides/multi-tenant) |
+| **Self-Hosted LLM Gateway** | Consume Pinecall's hosted open model (Qwen3) for chat and structured analysis over an authenticated, plan-gated streaming endpoint. | [`references/guides/self-hosted-llm.md`](references/guides/self-hosted-llm.md) · [docs](https://docs.pinecall.io/guides/self-hosted-llm) |
 | **SSE Event Streaming** | Stream agent events to your frontend in real time with Server-Sent Events. | [`references/guides/sse-streaming.md`](references/guides/sse-streaming.md) · [docs](https://docs.pinecall.io/guides/sse-streaming) |
 | **WebSocket Event Streaming** | Stream agent events over WebSocket for bidirectional, real-time communication with your frontend. | [`references/guides/ws-streaming.md`](references/guides/ws-streaming.md) · [docs](https://docs.pinecall.io/guides/ws-streaming) |
 | **Dev Mode** | Run dev and production agents on the same phone number, with zero extra Twilio cost. | [`references/guides/dev-mode.md`](references/guides/dev-mode.md) · [docs](https://docs.pinecall.io/guides/dev-mode) |
@@ -61,7 +62,9 @@ const agent = pc.agent("mara", {
 - **TTS model is auto-derived from `language`** — non-English agents (e.g.
   `language: "es"`) default ElevenLabs to `eleven_multilingual_v2` so numbers,
   dates and currency are pronounced correctly (flash/turbo don't normalize text).
-  English stays on `eleven_flash_v2_5`. Override with `voice: { ..., model: "..." }`.
+  English stays on `eleven_flash_v2_5`. To keep flash on a non-English agent
+  (lower latency/cost), set the top-level `flash: true` flag. To pin any model,
+  use `voice: { ..., model: "..." }` (explicit model always wins over `flash`).
 - **Greeting**: inbound → `greeting` field in `pc.agent()`; outbound → `greeting`
   field in `agent.dial()`. It is sugar for `call.say()` in `call.started`.
 - **Auth**: `new Pinecall()` reads `PINECALL_API_KEY` from env and auto-connects.

package/skills/pinecall-guides/references/guides/self-hosted-llm.md

@@ -0,0 +1,148 @@
+---
+title: "Self-Hosted LLM Gateway"
+description: "Consume Pinecall's hosted open model (Qwen3) for chat and structured analysis over an authenticated, plan-gated streaming endpoint."
+---
+
+# Self-Hosted LLM Gateway
+
+Pinecall hosts an open LLM and exposes it through an authenticated streaming
+endpoint on the sdk-server. Use it for any task that wants a cheap, in-house LLM
+instead of a paid per-token provider: **chat / agent loops** and **structured
+analysis** (classification, extraction, summarization, recommendations).
+
+| Model | Size | Best for |
+|-------|------|----------|
+| `qwen3:14b` | ~9 GB | **default** — hybrid model: clean JSON/analysis with thinking off, step-by-step reasoning with thinking on |
+| `deepseek-r1:14b` | ~9 GB | dedicated reasoning — **coming soon** |
+| `qwen2.5-coder:14b` | ~9 GB | code generation, refactors, tool/JSON authoring — **coming soon** |
+| `mistral-nemo:12b` | ~7 GB | strong multilingual + 128k context — **coming soon** |
+
+> Models flagged **coming soon** aren't live yet — `GET /api/llm/models` always
+> returns the currently available set.
+
+## Authentication & access
+
+- **Base URL:** `https://voice.pinecall.io`
+- **Auth:** a Pinecall API key via `X-API-Key: <key>` **or** `Authorization: Bearer <key>`.
+- **Plan gating:** **paid plans only** (`starter`, `pro`, `enterprise`). Both `free`
+  and `free_trial` receive **`402 SUBSCRIPTION_REQUIRED`**.
+
+## `POST /api/llm/chat`
+
+Streams the completion as **Server-Sent Events**.
+
+### Request body
+
+```jsonc
+{
+  "messages":    [{ "role": "user", "content": "..." }],  // required
+  "system":      "optional system prompt",
+  "model":       "qwen3:14b",                             // default: qwen3:14b
+  "mode":        "chat" | "analysis",                     // default: "chat"
+  "think":       false,                                   // reasoning on/off (default false; analysis forces false)
+  "temperature": 0.7,
+  "max_tokens":  512,
+  "format":      { /* JSON schema */ } | "json"           // analysis mode only
+}
+```
+
+Qwen3 is a **hybrid** model: `think: false` (the default) returns a clean, direct
+answer — best for JSON and low latency. `think: true` lets it reason step-by-step
+first (better on hard problems); the reasoning never leaks into the streamed
+answer. `mode: "analysis"` always forces thinking off so JSON stays clean.
+
+### SSE event stream
+
+```
+data: {"type":"token","content":"..."}                 // repeated — incremental text
+data: {"type":"done","usage":{"input_tokens":N,"output_tokens":M}}
+data: {"type":"error","error":"...","code":"UPSTREAM_ERROR|INTERNAL"}
+data: [DONE]                                            // terminator
+```
+
+### Errors
+
+| Status | Code | Meaning |
+|--------|------|---------|
+| 401 | `MISSING_KEY` / `INVALID_KEY` | no or bad API key |
+| 402 | `SUBSCRIPTION_REQUIRED` | tier is `free` or `free_trial` |
+| 400 | `MISSING_MESSAGES` / `BAD_MODEL` / `BAD_REQUEST` | invalid request |
+
+## `GET /api/llm/models`
+
+Same auth + gate. Returns the available models, the default, and the caller's tier —
+handy to probe access before streaming. **This is the source of truth for what's
+currently available** (the list grows over time).
+
+```json
+{ "models": ["qwen3:14b"], "default": "qwen3:14b", "tier": "pro" }
+```
+
+## Chat — streaming agent loop
+
+```ts
+const res = await fetch("https://voice.pinecall.io/api/llm/chat", {
+  method: "POST",
+  headers: {
+    "Content-Type": "application/json",
+    "X-API-Key": process.env.PINECALL_API_KEY!,
+  },
+  body: JSON.stringify({
+    model: "qwen3:14b",
+    system: "You are a concise assistant.",
+    messages: [{ role: "user", content: "Summarize today's bookings." }],
+    // think: true,  // ← opt into step-by-step reasoning for harder questions
+  }),
+});
+
+const reader = res.body!.getReader();
+const dec = new TextDecoder();
+let buf = "";
+for (;;) {
+  const { value, done } = await reader.read();
+  if (done) break;
+  buf += dec.decode(value, { stream: true });
+  for (const line of buf.split("\n\n")) {
+    if (!line.startsWith("data: ")) continue;
+    const data = line.slice(6);
+    if (data === "[DONE]") break;
+    const evt = JSON.parse(data);
+    if (evt.type === "token") process.stdout.write(evt.content);
+  }
+  buf = buf.slice(buf.lastIndexOf("\n\n") + 2);
+}
+```
+
+## Analysis — structured JSON (schema-enforced)
+
+Set `mode: "analysis"` and pass a JSON **schema** in `format`. The gateway routes
+analysis requests through a native path that constrains the output to your schema
+(and forces thinking off) — ideal for recommendations and extraction.
+
+```ts
+const body = {
+  model: "qwen3:14b",
+  mode: "analysis",
+  system: "You are a pricing engine. Return JSON only.",
+  messages: [{ role: "user", content: "Service: deep-tissue massage, $80, 95% utilization, 60% margin. Recommend an optimal price." }],
+  format: {
+    type: "object",
+    properties: {
+      suggestedPrice: { type: "number" },
+      confidence:     { type: "string", enum: ["low", "medium", "high"] },
+      rationale:      { type: "string" },
+    },
+    required: ["suggestedPrice", "confidence", "rationale"],
+  },
+};
+// POST as above, accumulate the `token` chunks into `text`, then:
+const rec = JSON.parse(text); // { suggestedPrice, confidence, rationale }
+```
+
+> **Warning:** Pass a real JSON-schema **object**. The string `"json"` (OpenAI-style
+> `response_format`) only nudges the model toward JSON — it does **not** enforce a shape.
+
+> **Note:** This open model is for **in-app responders, analysis, and
+> recommendations**. For **live voice / WhatsApp agents**, the Pinecall server-side
+> LLM supports OpenAI / Mistral / Google / Anthropic — see
+> [LLM Providers](/reference/llm-providers).

package/skills/pinecall-quickstart/SKILL.md

@@ -47,7 +47,9 @@ const agent = pc.agent("mara", {
 - **TTS model is auto-derived from `language`** — non-English agents (e.g.
   `language: "es"`) default ElevenLabs to `eleven_multilingual_v2` so numbers,
   dates and currency are pronounced correctly (flash/turbo don't normalize text).
-  English stays on `eleven_flash_v2_5`. Override with `voice: { ..., model: "..." }`.
+  English stays on `eleven_flash_v2_5`. To keep flash on a non-English agent
+  (lower latency/cost), set the top-level `flash: true` flag. To pin any model,
+  use `voice: { ..., model: "..." }` (explicit model always wins over `flash`).
 - **Greeting**: inbound → `greeting` field in `pc.agent()`; outbound → `greeting`
   field in `agent.dial()`. It is sugar for `call.say()` in `call.started`.
 - **Auth**: `new Pinecall()` reads `PINECALL_API_KEY` from env and auto-connects.

package/skills/pinecall-reference/SKILL.md

@@ -36,7 +36,9 @@ table below indexes every page; open the `references/…` file for the full text
 - **TTS model is auto-derived from `language`** — non-English agents (e.g.
   `language: "es"`) default ElevenLabs to `eleven_multilingual_v2` so numbers,
   dates and currency are pronounced correctly (flash/turbo don't normalize text).
-  English stays on `eleven_flash_v2_5`. Override with `voice: { ..., model: "..." }`.
+  English stays on `eleven_flash_v2_5`. To keep flash on a non-English agent
+  (lower latency/cost), set the top-level `flash: true` flag. To pin any model,
+  use `voice: { ..., model: "..." }` (explicit model always wins over `flash`).
 - **Greeting**: inbound → `greeting` field in `pc.agent()`; outbound → `greeting`
   field in `agent.dial()`. It is sugar for `call.say()` in `call.started`.
 - **Auth**: `new Pinecall()` reads `PINECALL_API_KEY` from env and auto-connects.

package/skills/pinecall-reference/references/reference/tts-providers.md

@@ -77,9 +77,46 @@ The server picks the ElevenLabs model from your `language`:
 | `en` (or unset) | `eleven_flash_v2_5` | Fastest, optimized for real-time streaming |
 | Any non-English (`es`, `fr`, `de`, …) | `eleven_multilingual_v2` | Flash/Turbo don't normalize text, so Spanish & other languages mispronounce numbers, dates, currency and abbreviations. The multilingual model reads them naturally. |
 
-> `eleven_multilingual_v2` is billed at a higher rate than flash (it's a higher-quality model). If you'd rather keep the faster/cheaper flash model for a non-English agent, pin it explicitly (see below).
+> `eleven_multilingual_v2` is billed at a higher rate than flash (it's a higher-quality model). If you'd rather keep the faster/cheaper flash model for a non-English agent, use the `flash` shortcut or pin the model explicitly (both below).
 
-**Override the model** with the optional `model` field — it always wins over the auto-default:
+#### `flash: true` — keep flash on a non-English agent
+
+The multilingual model trades a little **latency** for much better pronunciation.
+If your non-English agent should prioritize **lowest latency / lowest cost** over
+pronunciation quality, set the top-level `flash` flag — it opts out of the
+multilingual auto-default and keeps `eleven_flash_v2_5`:
+
+```typescript
+const agent = pc.agent("sofia", {
+  prompt: "Sos Sofía, asistente de la clínica.",
+  llm: "openai/gpt-5-chat-latest",
+  voice: "elevenlabs/agus",
+  stt: "deepgram/flux",
+  language: "es",
+  flash: true,        // ← stay on eleven_flash_v2_5 despite language: "es"
+});
+```
+
+`flash` is a sibling of `language` (not inside `voice`), so it reads cleanly with
+the rest of the shortcuts. Semantics:
+
+| Config | Resulting ElevenLabs model |
+|---|---|
+| `language: "es"` | `eleven_multilingual_v2` (auto) |
+| `language: "es"`, `flash: true` | `eleven_flash_v2_5` |
+| `language: "en"` (with or without `flash`) | `eleven_flash_v2_5` |
+| `voice: { model: "..." }` (any `flash`/`language`) | the pinned model — explicit always wins |
+
+Notes:
+
+- **ElevenLabs only.** `flash` has no effect on Cartesia or Polly.
+- **No-op for English** — English already defaults to flash.
+- **An explicit `voice: { model }` always wins** over `flash`. Use `flash: true`
+  for the common "I want the cheap fast model" case; use the `model` field when
+  you need a specific model id.
+- Works per-channel too: `phoneNumbers: [{ number, language: "es", flash: true }]`.
+
+**Override the model** with the optional `model` field — it always wins over both the auto-default and `flash`:
 
 ```typescript
 voice: {

package/skills/pinecall-sdk-api/SKILL.md

@@ -49,7 +49,9 @@ const agent = pc.agent("mara", {
 - **TTS model is auto-derived from `language`** — non-English agents (e.g.
   `language: "es"`) default ElevenLabs to `eleven_multilingual_v2` so numbers,
   dates and currency are pronounced correctly (flash/turbo don't normalize text).
-  English stays on `eleven_flash_v2_5`. Override with `voice: { ..., model: "..." }`.
+  English stays on `eleven_flash_v2_5`. To keep flash on a non-English agent
+  (lower latency/cost), set the top-level `flash: true` flag. To pin any model,
+  use `voice: { ..., model: "..." }` (explicit model always wins over `flash`).
 - **Greeting**: inbound → `greeting` field in `pc.agent()`; outbound → `greeting`
   field in `agent.dial()`. It is sugar for `call.say()` in `call.started`.
 - **Auth**: `new Pinecall()` reads `PINECALL_API_KEY` from env and auto-connects.

package/skills/pinecall-sdk-api/references/api/agent.md

@@ -35,7 +35,8 @@ const agent = pc.agent("my-agent", {
 | Config field | Type | Description |
 |---|---|---|
 | `voice` | `string \| VoiceConfig` | TTS provider — shortcut or full config |
-| `language` | `string` | BCP-47 language code |
+| `language` | `string` | BCP-47 language code. Non-English auto-selects ElevenLabs `eleven_multilingual_v2` |
+| `flash` | `boolean` | Keep ElevenLabs `eleven_flash_v2_5` on a non-English agent (lowest latency/cost) instead of the multilingual auto-default. ElevenLabs-only; see [TTS Providers](/reference/tts-providers) |
 | `stt` | `string \| STTConfig` | STT provider — shortcut or full config |
 | `llm` | `LLMConfig` | LLM provider, model, prompt, enabled flag |
 | `tools` | `Tool[]` | Declarative tools created with `tool()` + Zod schemas (auto-executed) |

package/skills/pinecall-security/SKILL.md

@@ -30,7 +30,9 @@ table below indexes every page; open the `references/…` file for the full text
 - **TTS model is auto-derived from `language`** — non-English agents (e.g.
   `language: "es"`) default ElevenLabs to `eleven_multilingual_v2` so numbers,
   dates and currency are pronounced correctly (flash/turbo don't normalize text).
-  English stays on `eleven_flash_v2_5`. Override with `voice: { ..., model: "..." }`.
+  English stays on `eleven_flash_v2_5`. To keep flash on a non-English agent
+  (lower latency/cost), set the top-level `flash: true` flag. To pin any model,
+  use `voice: { ..., model: "..." }` (explicit model always wins over `flash`).
 - **Greeting**: inbound → `greeting` field in `pc.agent()`; outbound → `greeting`
   field in `agent.dial()`. It is sugar for `call.say()` in `call.started`.
 - **Auth**: `new Pinecall()` reads `PINECALL_API_KEY` from env and auto-connects.

package/skills/pinecall-web-chat/SKILL.md

@@ -31,7 +31,9 @@ table below indexes every page; open the `references/…` file for the full text
 - **TTS model is auto-derived from `language`** — non-English agents (e.g.
   `language: "es"`) default ElevenLabs to `eleven_multilingual_v2` so numbers,
   dates and currency are pronounced correctly (flash/turbo don't normalize text).
-  English stays on `eleven_flash_v2_5`. Override with `voice: { ..., model: "..." }`.
+  English stays on `eleven_flash_v2_5`. To keep flash on a non-English agent
+  (lower latency/cost), set the top-level `flash: true` flag. To pin any model,
+  use `voice: { ..., model: "..." }` (explicit model always wins over `flash`).
 - **Greeting**: inbound → `greeting` field in `pc.agent()`; outbound → `greeting`
   field in `agent.dial()`. It is sugar for `call.say()` in `call.started`.
 - **Auth**: `new Pinecall()` reads `PINECALL_API_KEY` from env and auto-connects.

package/skills/pinecall-web-components/SKILL.md

@@ -30,7 +30,9 @@ table below indexes every page; open the `references/…` file for the full text
 - **TTS model is auto-derived from `language`** — non-English agents (e.g.
   `language: "es"`) default ElevenLabs to `eleven_multilingual_v2` so numbers,
   dates and currency are pronounced correctly (flash/turbo don't normalize text).
-  English stays on `eleven_flash_v2_5`. Override with `voice: { ..., model: "..." }`.
+  English stays on `eleven_flash_v2_5`. To keep flash on a non-English agent
+  (lower latency/cost), set the top-level `flash: true` flag. To pin any model,
+  use `voice: { ..., model: "..." }` (explicit model always wins over `flash`).
 - **Greeting**: inbound → `greeting` field in `pc.agent()`; outbound → `greeting`
   field in `agent.dial()`. It is sugar for `call.say()` in `call.started`.
 - **Auth**: `new Pinecall()` reads `PINECALL_API_KEY` from env and auto-connects.

package/skills/pinecall-web-voice/SKILL.md

@@ -33,7 +33,9 @@ table below indexes every page; open the `references/…` file for the full text
 - **TTS model is auto-derived from `language`** — non-English agents (e.g.
   `language: "es"`) default ElevenLabs to `eleven_multilingual_v2` so numbers,
   dates and currency are pronounced correctly (flash/turbo don't normalize text).
-  English stays on `eleven_flash_v2_5`. Override with `voice: { ..., model: "..." }`.
+  English stays on `eleven_flash_v2_5`. To keep flash on a non-English agent
+  (lower latency/cost), set the top-level `flash: true` flag. To pin any model,
+  use `voice: { ..., model: "..." }` (explicit model always wins over `flash`).
 - **Greeting**: inbound → `greeting` field in `pc.agent()`; outbound → `greeting`
   field in `agent.dial()`. It is sugar for `call.say()` in `call.started`.
 - **Auth**: `new Pinecall()` reads `PINECALL_API_KEY` from env and auto-connects.

package/skills/pinecall-web-widget/SKILL.md

@@ -34,7 +34,9 @@ table below indexes every page; open the `references/…` file for the full text
 - **TTS model is auto-derived from `language`** — non-English agents (e.g.
   `language: "es"`) default ElevenLabs to `eleven_multilingual_v2` so numbers,
   dates and currency are pronounced correctly (flash/turbo don't normalize text).
-  English stays on `eleven_flash_v2_5`. Override with `voice: { ..., model: "..." }`.
+  English stays on `eleven_flash_v2_5`. To keep flash on a non-English agent
+  (lower latency/cost), set the top-level `flash: true` flag. To pin any model,
+  use `voice: { ..., model: "..." }` (explicit model always wins over `flash`).
 - **Greeting**: inbound → `greeting` field in `pc.agent()`; outbound → `greeting`
   field in `agent.dial()`. It is sugar for `call.say()` in `call.started`.
 - **Auth**: `new Pinecall()` reads `PINECALL_API_KEY` from env and auto-connects.