LM-Kit OneDocs2026.11.1lm-kit.comEULA

Integration Recipes

An agent earns its keep when it acts in the systems your teams already use: it files the ticket, logs the call, sends the summary. A recipe sets that up in one step for a common system: the custom tools that call it, and an agent that works through them.

1What each recipe makes#

Recipe Makes Call Before each call
CRM Tool crm_find_customer GET {{base_url}}/customers Runs
CRM Tool crm_log_activity POST {{base_url}}/customers/{customer_id}/activities Asks for approval
CRM Agent draft crm-assistant Works through crm_find_customer, crm_log_activity Its tools' own rule
Email summaries Agent draft email-summary Works through smtp_send Its tools' own rule
Helpdesk Tool helpdesk_get_ticket GET {{base_url}}/tickets/{ticket_id} Runs
Helpdesk Tool helpdesk_create_ticket POST {{base_url}}/tickets Asks for approval
Helpdesk Tool helpdesk_add_note POST {{base_url}}/tickets/{ticket_id}/notes Asks for approval
Helpdesk Tool helpdesk_escalate_ticket POST {{base_url}}/tickets/{ticket_id}/escalate Asks for approval
Helpdesk Agent draft helpdesk-support, from the customer-support blueprint Works through helpdesk_get_ticket, helpdesk_create_ticket, helpdesk_add_note, helpdesk_escalate_ticket Its tools' own rule

A tool that writes asks for approval before each call: the run stops in Approvals with the arguments it would send, and continues once a person approves. A tool that only reads runs.

2Install one#

Open Admin > Build > Custom tools and choose Install beside the recipe. Enter the system's API address and the credential it expects in the Authorization header (a bearer token, or Basic and the encoded key), check what the recipe makes, and choose Install.

Installing:

  • makes the recipe's tools as custom tools, the credential sealed in the store and never shown again;
  • makes its agent as a draft: review it, choose its knowledge sources, run its starter evaluation and publish it in Agents; nobody is served the agent until then;
  • replaces nothing: when a tool or an agent already has one of the names, the install is refused and names it;
  • changes no setting: a server tool the agent needs (smtp_send for email) is named, and stays off until you turn it on in Tools with the mail settings it needs.

Each tool stays yours to adapt: the recipes call a common REST shape (/tickets, /customers/{customer_id}/activities), and Edit on the Custom tools page maps a tool to your system's own paths and fields.

3Helpdesk: tickets with approvals#

The helpdesk agent starts from the customer support blueprint: it answers from your help center, and when a person must take over it files a ticket, adds what the customer said to an existing one, and escalates legal and safety matters.

  1. Install the recipe, open the helpdesk-support draft in Agents, choose its help-center collections, run its evaluation and publish it.
  2. Callers that start it as a run (POST /lmkit/v1/agents/helpdesk-support/runs) get a run that stops at each ticket call; whoever holds the approvals inbox approves it, and the run carries on with the ticket id.
  3. A webhook on run.awaiting_approval tells the team a ticket waits for a decision without anyone watching the inbox (see Webhooks).

4Credentials for connectors#

The same rule holds for MCP connectors (Admin > Build > Connectors): Credentials on a saved connector sets the bearer token and headers an http connector sends, or the environment variables a local command starts with. They are sealed, never read back (the panel shows which are set and when), replaced at any time, and never carried by a settings export; the connection reconnects with them at once.

5In the SDK#

The recipes are plain JSON definitions, and the same definitions run in process: the HTTP API recipes sample loads a recipe, fills in your address and credential, and gives its tools to an agent with HttpApiTool.

  • Custom Tools: every field of a tool definition.
  • Agent Blueprints: the agents the recipes start from.
  • Agent Runs: approvals, and runs that wait for them.