Skip to main content

Use Cases

The Bolti Model Context Protocol (MCP) server shines when you describe an outcome in plain language and let your client's model orchestrate the tool calls. Below are common workflows, each with an example chat prompt and the tool sequence it typically triggers.

These assume you have the server installed and authenticated - see Installation. Workspace and organization are auto-resolved from your session, so you rarely need to name them.

Confirmation

Any step that is destructive or billed (marked below) pauses for a confirm: true before it runs. Your client asks you first.

1. Review recent calls and pull a transcript​

Read-only investigation is the fastest win. Bring call data straight into your chat for analysis.

"Show me the last 10 calls for the support agent, then pull the transcript of the longest one."

1. agents_list                     -> find the support agent's ID
2. calls_list                      -> last 10 calls for that agent
3. agent_conversations_list        -> map calls to conversations
4. agent_conversation_get          -> transcript of the longest conversation

You can follow up with call_recording_url_get to get a signed audio playback link.

2. Iterate on an agent's system prompt​

Editing prompts in chat is a tight feedback loop: read, change, save.

"Show me the support agent's system prompt, then make the greeting warmer and save it."

1. agent_get      -> read the current `prompt` field
2. agent_update   -> write the revised `prompt` back

Because the prompt supports Liquid templating, you can introduce a variable at the same time, for example changing a hard-coded company name to {{ company_name | default: "Acme" }}.

3. Create a workspace tool and attach it to an agent​

Give an agent a new capability end to end without leaving chat.

"Create a tool called lookup_order that does a GET on https://api.acme.com/orders/{{ order_id }} with an order_id string argument, test it with order ORD-9182, then assign it to my support agent."

1. tool_create        -> URL + method + input_schema declaring `order_id`
2. tool_test          -> run against ORD-9182 with mock arguments
3. agent_tools_list   -> read the support agent's current tools
4. agent_tools_set    -> full list of tool UUIDs including the new one

The tool's input_schema (a JSON Schema) defines what the agent's LLM must supply; reference those arguments in the URL, headers, or body with {{ argument }}. See Workspace HTTP Tools.

4. Place an outbound call with per-call variables​

Prompts with variables let one agent serve many callers. Check what a call needs, then place it.

"Place a call from the sales agent to +14155551212 for customer Priya about her trial ending Friday."

1. agent_variables_get   -> see which prompt variables are required
2. outbound_call_start   -> (confirm) pass variable_values:
                            { customer_name: "Priya", topic: "trial ending Friday" }

outbound_call_start is billed, so it requires confirm: true. Running agent_variables_get first prevents a failed call from a missing required variable.

5. Provision a phone number for inbound calls​

Wire up inbound calling from your provider's available numbers.

"Find an available US number and assign it to the support agent for inbound calls."

1. dids_list_available   -> list unassigned Direct Inward Dialing numbers
2. phone_number_assign   -> attach a chosen number to the support agent
3. phone_numbers_list    -> confirm the assignment

6. Audit workspace configuration and credits​

Answer operational questions in one turn instead of clicking through the dashboard.

"Who has access to this workspace, which models are we using, and how many credits are left?"

1. workspace_members_list   -> members and their roles
2. supported_models_list    -> models available to the workspace
3. workspace_credits_get    -> current credit balance

For a workspace-wide cleanup, you can chain reads across agents - for example agents_list then agent_tools_list per agent - to report which tools are assigned where, then fix outliers with agent_tools_set.

Tips for good results​

Get more reliable runs
  • Name the agent, not its UUID. The model resolves IDs from agents_list; you do not need to remember them.
  • Ask read-first. When exploring, request the read tools explicitly so you can eyeball data before any write.
  • Expect a confirmation prompt on destructive or billed steps - that pause is by design.
  • Run agent_variables_get before an outbound call so you supply every required variable.

Where to go next​

If you want to...Read
See every tool by nameAvailable Tools
Understand token scope and safetyPersonal Access Tokens
Learn the tool-authoring model in depthWorkspace HTTP Tools