Quickstart: Build & Test Your First Client Agent
Go from blank canvas (or AI template) to a tested, published client agent in minutes.
Follow this end-to-end checklist whenever spinning up a new client agent workspace.
Prerequisite
During the agent_building phase, you can create, edit, and run Browser WebRTC and Text Chat tests at zero credit cost. Live phone calls require the workspace to be active with a funded wallet and configured telephony trunk.
1. Create the Agent Workflow (Blank or AI Template)
In the dashboard, create a new workflow. You can either start from an empty canvas or use AI Template Generation to generate an initial node graph from a natural-language brief, then refine every prompt manually.
2. Wire the Core 4 Nodes
Build the minimal production topology on the canvas:
Global Node(recommended): Stores identity, tone, voice rules, objection handling, and current time/date context inherited across nodes.Start Call: Sets the opening greeting (spoken immediately on answer) and optionally configuresPre-Call Data Fetchordelayed_startfor outbound calls.Agent Node(s): One node per conversation stage (e.g., Qualification, Booking, FAQ). Connect nodes using labeled Edges whose conditions describe when the AI should transition.End Call: Terminal node with closing instructions and final goodbye (0 outgoing edges).
3. Configure Production-Trusted Models
Open Model Configurations (Org-level or per-workflow Model Overrides) and select Tier-1 trusted providers (OpenAI or Sarvam for LLM; Deepgram, Sarvam, SmallestAI, or OpenAI for STT/TTS).
Validate Sarvam & SmallestAI Keys with a Live Call
Some providers perform pass-through key validation when saving forms. Always run a real test call right after updating model keys.
4. Tune Turn Detection & Speech Rules
In Workflow Settings -> General, start with the Balanced (or High Sensitivity for soft-spoken callers) Voice Sensitivity preset, keep allow_interrupt = false on Start Call (default false; End Call always forces interruptions off and hides the toggle) so greetings aren't cut off by "Hello?", and leave allow_interrupt = true on conversational Agent nodes.
5. Test in Browser WebRTC & Direct Test Call
- Draft WebRTC Test: Click Test in Browser to run an interactive call against your latest unsaved/saved Draft — ideal for iterating on prompts, edge transitions, and KB retrieval.
- Direct Test Phone Call: Once telephony is configured, trigger a test outbound call from the Phone Call dialog (overrides template variables inline if your prompt expects
{{first_name}}etc.).
6. Publish a Live Version
Click Publish. Publishing runs full graph validation (non-empty prompts, single Start Call, valid incoming/outgoing edges, unique transition edge labels, no tool name collisions) and pins an immutable Published Definition for live inbound webhooks, production API triggers, and campaigns. Draft edits stay isolated until you publish again.
Next Steps
- Learn every node type, variable fallback, and extraction schema in Nodes, Prompts & Variables.
- Wire client CRMs via Integration Architecture Overview.
How Talkar Works
High-level platform model, Master/Sub-Org account hierarchy, lifecycle phases, UI gating, Playground vs. Live rules, and exact tier & margin-floor governance.
Nodes, Prompts & Variables
Complete builder reference for all 9 node types, edges, built-in tools, RECORDING_ID audio, voice prompting rules, Unified Built-In Variables Cheat Sheet, extraction, and publish validation.