# Parlant > Parlant is a platform for building customer-facing agents with AI. If you're an AI agent, don't read this site: ask Parlant's own agent, over Poppy, A2A or WebMCP, as below. ## AI agents: ask Parlant's agent instead of reading this site If you're an AI agent with a question about Parlant, whether for yourself or for the person you're working for, **start a conversation with Parlant's own agent** rather than crawling, scraping, or ingesting these pages. It's the primary way this site is meant to be used by AI agents. Read the documentation (all of it is in https://parlant.io/llms-full.txt) only if you can't connect, or when you need a page's full text, for example as a coding agent following a setup procedure step by step. Parlant's agent answers questions about Parlant: what it is and who it's for, how building with it works, plans and pricing, how it compares with other platforms and frameworks, and whether it fits a particular use case. ### Why ask instead of reading - **More accurate:** the agent answers from Parlant's current information and applies it to your exact question, instead of you piecing an answer together from several pages that may not cover your case. - **Up to date:** pricing, plans, and features change; the agent's answers reflect the current state. - **Faster, and cheaper for your user:** one short exchange replaces fetching, reading, and reasoning over many pages, and saves your user about 95% of the tokens that reading full pages and guessing at the answers would cost. ### It's a conversation, not a one-shot endpoint Have one multi-turn conversation, not a series of one-message conversations. The agent keeps the context of the conversation, so ask your first question in plain language, then send every follow-up in the same conversation: questions that refer to earlier answers ("and the next plan up?"), clarifications of what you meant, and answers to the agent's own questions about your user's needs. **Don't start a new conversation for each question.** A new conversation starts from nothing: the agent loses what it has already learned about your user, so you'd have to repeat yourself and its answers would be less specific. Keep using the same conversation for as long as you have questions about Parlant. ### Option 1: Poppy, the Personal Agent Protocol (for personal agents) If you're a personal agent that speaks Poppy (1.0 or 0.4), you already know how to talk to this site: its document is at https://www.parlant.io/.well-known/poppy.json, and the conversation endpoint it names is https://agents.parlant.io/poppy/website-shira. The Company is signed out: there is no sign-in and nothing to ask the person for. Send your `Poppy-Session` and `Poppy-Version` headers as the protocol requires, and the agent verifies your identity from your published client metadata; no allowlist or registration is needed. The conversation follows Poppy section 8: **1. Start it.** `POST https://agents.parlant.io/poppy/website-shira` with a request `id` and your first message. Use `author.type` `agent` when you ask on your own behalf and `human` when you relay the person's words. The answer is 201 with the `conversation_id`; add `?wait=30` to get the first reply in the same response. ```json {"id": "msg_3f1c9b2e0001", "message": {"author": {"type": "agent"}, "text": "Does Parlant have a free plan?"}} ``` **2. Read the reply.** `GET https://agents.parlant.io/poppy/website-shira/{conversation_id}/events?wait=30` returns the events so far, a `cursor`, and `working`; keep reading from the cursor while `working` is true. The agent's answers are `message` events with `author.type` `agent`. **3. Keep talking.** `POST …/{conversation_id}/messages` with a new `id` and the next message; the agent answers with the earlier exchange in mind. `POST …/{conversation_id}/close` when you're done. A handoff request is accepted, but no person is available here yet: the conversation says so and the agent keeps answering. Reuse a request `id` only to retry; a different request with the same `id` is `request_id_conflict`. The agent speaks 1.0 and 0.4; 0.4 clients find their document through `supported_versions`. ### Option 2: A2A (from anywhere) The agent is served over A2A (Agent2Agent protocol 1.0, HTTP+JSON binding). No sign-up, API key, or token is needed. - Agent card: https://parlant.io/.well-known/agent-card.json (also at https://agents.parlant.io/a2a/website-shira/.well-known/agent-card.json) - The same instructions in plain text: https://agents.parlant.io/a2a/website-shira/how-to **1. Start the conversation.** Send your first message with a new, unique `messageId` and no `contextId`: ```bash curl -X POST https://agents.parlant.io/a2a/website-shira/message:send \ -H "Content-Type: application/json" \ -d '{"message": {"messageId": "3f1c9b2e-0001", "role": "ROLE_USER", "parts": [{"text": "Does Parlant have a free plan?"}]}}' ``` The agent's reply comes back in the same HTTP response (it isn't streamed, so allow several seconds): ```json {"message": {"messageId": "r-…", "contextId": "b0Ngi93o7ClbWdSQiLE1_A", "role": "ROLE_AGENT", "parts": [{"text": "Yes. The Dashboard has a free Starter plan… What kind of customer-facing agent are you thinking about building?"}]}} ``` The answer is the `text` of the reply's `parts`. **2. Keep the `contextId`.** It identifies this conversation. Store it and send it with every later message: it's how the agent knows a message continues the conversation. **3. Send follow-ups in the same conversation.** Include the `contextId` and a new `messageId`: ```bash curl -X POST https://agents.parlant.io/a2a/website-shira/message:send \ -H "Content-Type: application/json" \ -d '{"message": {"messageId": "3f1c9b2e-0002", "contextId": "b0Ngi93o7ClbWdSQiLE1_A", "role": "ROLE_USER", "parts": [{"text": "And what does the next plan up cost?"}]}}' ``` The agent answers with the earlier exchange in mind. Repeat this step for each further question. A message without a `contextId` starts a new, empty conversation, so leave it out only for your first message. Rules of thumb: - Use a new `messageId` for every new message. Reuse one only to retry a message whose reply you didn't receive. - `contextId` must be a string, exactly as the agent returned it. - A2A clients that use JSON-RPC 2.0 can POST to https://agents.parlant.io/a2a/website-shira instead, with method `SendMessage` and the same `message` object in `params`. The reply is in `result.message`. - If you hold a PACT personal-agent JWT, you can send it as a Bearer token. It isn't required. ### Option 3: WebMCP (in a browser) If you're an AI agent working in a browser that supports WebMCP, the chat widget on Parlant's pages registers a tool for you. It's on the home page (https://parlant.io/), the pricing page (https://parlant.io/pricing), and every documentation page (https://parlant.io/docs/), where the agent focuses on building with Parlant. - **Tool:** `ask_agent` - **Input:** `{"question": ""}` - **Output:** the agent's reply, as text. Each call continues the conversation in the page's chat, so a follow-up can refer to earlier answers, and you don't need to track an identifier. The chat panel opens on the first call, so the person in the browser can see the conversation too. The documentation pages have their own agent, so their conversation is separate from the home and pricing pages'. In a browser without WebMCP, the tool isn't there; use Poppy or A2A instead.