---
title: "Start Agent Session Voice"
method: POST
path: "/agent/sessions/{session_id}/voice"
tags: ["Agents"]
---

# Start Agent Session Voice

`POST /agent/sessions/{session_id}/voice`

## Request

- Base URL: `https://be.graph8.com/api/v1`
- URL: `https://be.graph8.com/api/v1/agent/sessions/{session_id}/voice`
- Auth: HTTP bearer

Start speaking with the agent. THIS SPENDS CREDITS AS IT RUNS.

A live metered session: it begins charging when it connects and keeps
charging until it is ended or hits `max_credits`. Quote it first.

`max_credits` is a HARD CEILING and is required - there is no default,
because choosing one would mean this contract picking a spending limit on
the customer's behalf. `rate_version` must be the one the quote returned.

End it with `DELETE /agent/sessions/{session_id}/voice/{voice_id}`. A
session left open keeps costing money.

## Path parameters

- `session_id` string, required: Session id (UUID).

## Headers

- `Authorization` string, nullable
- `X-Target-Org-Id` string, nullable

## Request body

- VoiceSessionRequest: Mirror of `VoiceSessionRequest`, including its `extra="forbid"`.
  - `voice_id` string, uuid, required: Which voice to speak with.
  - `rate_version` string, required: The rate version the quote was issued against, from `POST /agent/sessions/{session_id}/voice/quote`. The server refuses a stale one rather than charging today's price for yesterday's quote.
  - `max_credits` integer, required: HARD CEILING on what this session may spend. Required, with no default: an uncapped voice session is an open-ended charge, and a contract that defaults it would be choosing a spending limit on the customer's behalf.

## Response `200`

Successful Response

- ApiResponseAgentSessionStatePayload
  - `data` AgentSessionStatePayload, required: An open envelope for a hydration or voice answer. Open because session hydration returns the agent's STEPS as well as its messages, and a step's `detail` is whatever the tool that produced it reported. That set grows with every tool the platform gains, and a pinned mirror would drop the detail of exactly the newest ones - which on a transcript surface reads as the agent having done less than it did.
  - `pagination` PaginationMeta
    - `page` integer, required: Current page number (1-indexed)
    - `limit` integer, required: Items per page
    - `total` integer, required: Total number of items
    - `has_next` boolean, required: Whether there are more pages
    - `next_cursor` string, nullable: Opaque cursor for the next page; null when there are no more items. Pass it back as the `cursor` query param to fetch the following page without computing page numbers.
    - `total_available` integer, nullable: Search endpoints only: how many records really match, when known. Can be larger than `total`, which is capped at the 10,000 results that can be paged. For company search this is a distinct-company estimate. Null when unknown; absent outside search.
    - `total_is_lower_bound` boolean, nullable: Search endpoints only: true when more records match than `total` shows (the result set is larger than the 10,000 results that can be paged). Absent outside search.
    - `notes` string[], nullable: Search endpoints only: plain-language notes about the result, e.g. that more records match than can be paged, or that a reversed `between` range was put in order.

## Other responses

- `422`: Validation Error

---

[API](https://skmtc.dev/graph8/apis/graph8-developer-api.md) · [All operations](https://skmtc.dev/graph8/apis/graph8-developer-api/llms.txt) · [OpenAPI document](https://skmtc.dev/graph8/apis/graph8-developer-api/revisions/2e7cfa0ca921?raw)