Deploy a call-center agent
with tools and guardrails
This demo shows a voice support agent for a fictional carrier, built on Claude through the Chalk Model Gateway. Talk to an AI agent with real tools and real limits. It verifies you, fixes your bill or shuts off a stolen phone, and knows when to hand it off. Hang up and it files its own call notes and QA scorecard.
Call the agent
Pick a caller and press Call to talk, or watch to hear a simulated caller run the scenario: a double charge, a stolen phone, a cancellation threat, and an upcoming trip abroad. Then hang up and read the QA scorecard.
Try this experiment · expected results
1Start with Double charged
Use Call to play the customer, or Watch for the simulated caller.
Expected: The agent verifies the fictional caller before showing account details.
2Inspect an account action
Ask for a $200 credit and compare the reply with the tool result.
Expected: The example’s tool enforces its $50 per-call limit.
3Review the call
Select Hang up after the conversation.
Expected: Call notes and a QA scorecard cite the conversation and tool results.
Choose Watch to hear a simulated call without a microphone. Follow verification, tool actions, and the after-call summary; model calls go through Chalk’s gateway.
Halyard Mobile and its customers are fictional. Voice input uses your browser's speech recognition (Chrome, Edge or Safari); you can always type instead.
Built on Chalk
Chalk hosts the application and routes its model calls. The example supplies the fictional carrier, per-call account data, verification rules, guarded tools, and post-call QA rubric.
Route the model calls
The support agent and reviewer use the Chalk Model Gateway. The gateway manages provider credentials, usage, and traces; the application uses its configured gateway key.
Run on a Chalk scaling group
The example runs in a Chalk scaling group on a CPU replica. Its models are accessed through the gateway rather than running on that replica.
Enforce rules in the example’s tools
The application requires verification before account access, caps credits at $50 per call, and limits retention offers. These rules are implemented by the example, not inferred from the model’s prose.
Review the transcript and tool results
After the call, the example asks a second model for notes and a QA scorecard. Inspect the cited transcript and tool outcomes to assess the review.
Deploy your own in four commands
Clone the example, log in, create a Chalk Model Gateway key, and run the deploy script. Chalk builds the image, starts one small replica, and prints a URL that serves this exact demo. It is up in about a minute.
Replace the fake CRM with your own systems, write the policies into the tools, and add the checks your QA team already scores by hand.
git clone https://github.com/chalk-ai/call-center-agent
cd call-center-agent
# Or export CHALK_CLIENT_ID and CHALK_CLIENT_SECRET.
chalk login
# The agent reaches Claude through the Chalk Model Gateway with this key.
chalk router api-key create # store it as the secret CHALK_ROUTER_API_KEY
./deploy.pyPut an agent on your support line
Voice, chat or email, with your systems as the tools and your policies as the guardrails. Talk to an engineer about running support agents on Chalk.