Agent
One object,config, plus what you attach to it: a prompt, a model stack, tools, knowledge bases and handoff destinations. It runs on exactly one channel. → agents overview
Agent task
An agent nobody talks to: it takes named inputs, does the work and returns typed output fields. No conversation, no publish step — saving is live. → tasks overviewAgent team
Several agent definitions sent on one call request, the first of which answers and the rest of which are reachable by handoff. Built per request rather than stored. → agent teams on a callAvatar
A rendered face and voice for avideo agent, supplied by Anam. Billed by wall-clock minute, idle time included. → video avatars
Batch
A list of recipients worked through on a policy: calling windows, concurrency, retries, and pause, resume and cancel. Call batches dial; email batches draft and send. → batch callingBYOK
Bring your own key. Talqing holds no provider credentials of its own — every model runs on your account, on your key, billed to you by that provider. → provider keysCall
One voice or video session: an inbound phone call, an outbound one, or a call in a browser. The record you read a transcript, latency breakdown and cost from. → callsCascade
The three-stage voice pipeline: speech-to-text, then a language model, then text-to-speech. Three swappable parts, as opposed to one realtime model. → modelsChannel
voice, video or text. An agent has exactly one, and it decides which model slots exist. → channels
Close reason
The specific string saying how a call ended — the caller hung up, the agent ended it, it timed out, it crashed. Some reasons waive the platform fee. → close reasonsContact
The person, across every conversation you have had with them, and theuserdata accumulated about them. → userdata
Contact key
Your own stable string identifying a contact, unique in the workspace. Send the same key again and you reach the same thread and the same accumulated state. → text conversationsConversation
The thread with one person on one surface — a web chat, a Telegram thread, a single phone call. It can contain several sessions. → conversationsCredits
A prepaid USD balance, one per region, that the platform fee is debited from. Nothing else is charged against it. → pricing and creditsDraft
The version of an agent or tool every write lands on. Drafts run nothing and reach nobody until you publish. → versionsEndpointing
Deciding that a caller has finished speaking, so the agent may answer. The window is tunable and is the main lever on how interruptible an agent feels. → speech-to-text and turn detectionHandoff
Passing a conversation from one agent to another on Talqing, with the platform still running it. Each entry inhandoffs becomes a tool the model can call. Not a transfer. → handoffs
Integration
A connection to an outside service: an MCP server whose tools an agent may call, or a channel such as Telegram that an agent answers on. → integrations overviewItem
One entry in a conversation: a message from either side, a tool call, or a tool result. What a transcript is made of. → conversationsKnowledge base
Content crawled from your own URLs, indexed so an agent can answer from it. Only aready knowledge base can be attached to an agent. → knowledge overview
MCP
Model Context Protocol. Two directions: an MCP server you attach so your agent can call its tools, and Talqing’s own MCP server, which gives a coding agent one tool per API endpoint. → MCP servers and our MCP serverOperation
One step in a tool’s ordered tree — call an API, run TypeScript, branch, speak, hand off, transfer, end the call. Eleven kinds. Never called a step or a node. → operation treePersonal access token
Talqing’s own credential for the API, sent asAuthorization: Bearer <token>. It authenticates as its creator, carries their live role, and has no prefix. → authentication
Platform fee
The only money Talqing charges: 0.01 per minute of video, nothing on text. Waived on a failed call. → pricing and creditsPriority lane
llm.priority, a premium lower-latency lane some models offer, billed at that provider’s higher priority rates. → priority lane
Provider key
Your OpenAI, Deepgram, ElevenLabs or Anam key, stored write-only in the workspace and used to run your agents. Not a Talqing credential. → provider keysPublished
The immutable snapshot live traffic runs. Publishing freezes the current draft as the next version and pins every attached tool to the version live at that moment. → versionsRealtime
One speech-to-speech model in place of the cascade’s three stages. Lower latency, less control: scripted lines become instructions the model may reword. → realtimeRegion
Where a workspace’s resources physically live, chosen purely by base URL —https://api.in.talqing.com or https://api.us.talqing.com. No field in the API carries one. → regions
Secret
A workspace credential stored write-only and read as{{secrets.NAME}} by tools and integrations. The model never sees the value. → secrets
Session
One agent’s participation in a conversation — one phone call, one text window, one trigger firing. Usage, cost, latency and the close reason are per session. → conversationsSystem variable
{{system_vars.…}}: six read-only keys the platform supplies — human_phone_number, agent_phone_number, direction, now, date and time. A closed set. → variables
Tool
One capability the model can call mid-conversation: a name, a description, ajson_schema of arguments, and an ordered tree of operations. Published separately from the agent. → tools overview
Transfer
Handing a phone call to a human being on a real number, after which Talqing steps out. The destination is a literal number fixed at publish. Not a handoff. → transferTrigger
A binding from an inbound event on an integration — a Telegram message, say — to the text agent that should answer it. → triggersUserdata
{{userdata.field}}: what you know about the person. Seeded when a session starts, writable by any tool, and merged onto the contact record so it is there next time. Visible to the model. → userdata
Variable (vars)
{{vars.name}}: configuration of the session, declared on the agent with a description and a default, supplied by the request that starts the session. Nothing inside the session can write it. Visible to the model — never put a credential here. → variables
Workspace
The boundary everything belongs to: agents, tools, keys, numbers, calls and the people who can see them. Members holdADMIN, EDITOR or VIEWER. Called an organization in the dashboard. → workspaces and roles