Talqing is where you build an AI agent, publish it, put it in front of real people, and read back every conversation it had. An agent is a system prompt, a model stack, the tools it may call and the knowledge it can look things up in. Anything you could have written by hand as a LiveKit agent, you can build here.

​
The three channels

An agent runs on exactly one channel, and the channel decides which model slots it has. There is also the agent task — an agent nobody talks to, which takes named inputs, does some work and returns a typed result. Use one for research, classification and drafting, where there is an input and an answer rather than a conversation.

​
Strictly bring your own keys

Talqing holds no provider credentials of its own. You connect your own OpenAI, Deepgram, ElevenLabs, Gemini, xAI, Sarvam, Soniox, Raya, Anam or ai-coustics accounts, those providers bill you directly at your own rates, and none of that spend passes through us. What Talqing charges is a per-minute platform fee, and only on the channels that carry media: 0.0035perminuteofvoice,0.0035 per minute of voice, 0.01 per minute of video, and nothing at all on text. A text conversation costs you only what your own language-model key was charged. See pricing and credits. The consequence to plan around: nothing runs until you have added a key for every provider your agent uses. Publishing an agent whose provider has no key is refused, naming the provider. See provider keys.

​
Four ways to build the same thing

Every one of these works from the same operation surface — one operation per API endpoint — so nothing is reachable in one and missing from another.
  • The dashboard at https://app.talqing.com — editors for agents, tools, knowledge bases and tasks, plus telephony, conversations and observability.
  • The API and the SDKs — /v1 over HTTPS, with a TypeScript and a Python client generated from the same document.
  • The CoPilots — a build assistant inside the agent, tool, task and knowledge editors. It calls the same operations you would, as you, with your role.
  • Your own coding agent — point Claude Code, Codex or any MCP client at our MCP server and it gets one tool per endpoint plus the same platform knowledge the CoPilots work from.

​
What you get

Agents with drafts and published versions, version history and rollback. Tools built as an ordered tree of operations — call an API, run TypeScript, branch, speak, hand off, transfer to a human, end the call. Knowledge bases crawled from your own site. Agent tasks with typed output. Phone numbers imported from your own carrier account, answering and placing calls. Batch calling with scheduling windows, retries and pause/resume. Integrations — external MCP servers, native Google Calendar tools, Telegram. Observability — every transcript, tool call, latency stage and cost line, plus webhooks.
Everything a workspace owns is regional. Agents, tools, calls, phone numbers and the credit balance belong to the base URL they were created against — https://api.in.talqing.com or https://api.us.talqing.com. One token reaches every region; it names your workspace, not a place. See regions.

​
Start here

Build a voice agent

Dashboard only. Sign in, add keys, publish an agent and talk to it in the browser.

Take a real phone call

Connect your carrier, import a number, point an agent at it.

Talk to a text agent in code

One script, end to end, in TypeScript or Python.

Build from your terminal

Point Claude Code or Codex at the MCP server and build by asking.

Core concepts

The vocabulary, in the order you meet it.

Provider keys

Which providers, how to add a key, what fails without one.