Router agents, oversight, API access, and more
Everything that isn't in the day-one workflow: multi-agent routing, quality oversight, programmatic access, MCP, and custom AI providers. Skim what you need — nothing here is required to run the platform.
#Router agents
A router is a special agent whose job is to dispatch— read the customer's first message, pick the specialist agent that should handle the conversation, and hand off. The specialist then owns the conversation from that point forward (sticky routing).
When to use
- Multi-product businesses where different agents know different catalogues.
- Split Sales / Support agents so each has a focused system prompt.
- Multi-language teams — a router picks the language-matched specialist on the first message.
Setup
Create the specialist agents
Regular agents — one per intent / language / product. Fill in short Router description fields on each; the router reads these to decide who to pick.
Create the router agent
Go to /dashboard/ai/routers → New router. Router agents have isRouter=true — this unlocks the handoffTo and unassignAgent tools.
Pick which specialists it can dispatch to
The Routable agents panel lets you allow-list which specialists this router can pick. Empty list = every workspace sibling is fair game.
Bind the router to a channel
On the WhatsApp/Instagram instance card, set Router agent instead of Primary agent. The router handles the first turn, calls handoffTo(agentId), and the picked specialist takes over.
Once a contact is assigned to a specialist, they stay with that specialist even after the router agent is retired or reconfigured. Use unassignAgent from any specialist to send them back through the router.
#Oversight agents
An oversight agent reviews a sample of every regular agent's replies and flags issues you'd have missed — awkward phrasing, off-topic answers, missed handoff opportunities. Think of it as a QA teammate that reads every conversation.
Setup
- Go to
/dashboard/oversight. - Create an oversight agent — pick a model (any provider), describe what you're looking for.
- Point it at one or more of your regular agents.
Reviewing suggestions
Oversight suggestions land in a queue on the same page. You can Approve a suggestion — the fix is auto-applied to the reviewed agent's system prompt — or Reject to dismiss.
#API Keys
Every workspace can mint API keys for programmatic access — perfect for internal automation, custom UIs, or piping data into a warehouse.
Scope
Keys inherit the creator's workspace access. Rotate or revoke from /dashboard/api-keys.
Using a key
Include it as a bearer token in the Authorization header:
curl -H "Authorization: Bearer sk-... " \
https://chatbot.tural.ai/api/agents#Webhooks
Subscribe to workspace events and receive them on a URL you control. Useful when you want to pipe conversations into an external analytics tool, sync contacts to your own CRM, or trigger downstream workflows.
Event types
message.new— a new inbound or outbound message.client.updated— a CRM contact was created or changed.automation.executed— an automation ran.
Setup
Add a webhook at /dashboard/webhooks. Pick the events you want and a URL. Payloads are POSTed as JSON; retries with exponential backoff on non-2xx responses.
#MCP (Model Context Protocol)
Expose your alChatBot workspace to LLM clients like Claude Desktop via MCP. Once connected, Claude can search contacts, read conversations, and even send messages on your behalf.
Connect Claude Desktop
Go to /dashboard/mcp. Copy the MCP server URL, then in Claude Desktop go to Settings → Connectors → Add MCP server and paste it. You'll be prompted to authorise with your alChatBot account.
MCP is designed for personal use — reviewing conversations, extracting insights, drafting replies from your desktop. For production automation, use API keys + webhooks instead.
#Custom AI providers
The default AI providers cover most cases — but if you need Azure OpenAI, a self-hosted LLM, or a niche provider with an OpenAI-compatible API, you can add it as a custom provider.
Adding a custom provider
On /dashboard/ai/providers, click Add custom. You'll be asked for:
- Name — displayed everywhere else.
- Base URL — the OpenAI-compatible endpoint (e.g.
https://your-azure-openai.openai.azure.com/openai/deployments/gpt-5). - API Key — bearer token.
- Model list — which model ids to expose in the agent dropdown.
Custom providers must implement the OpenAI Chat Completions API surface, including tool-calling. Providers that don't support tool calls will crash on any agent with skills enabled.