Skip to main content
A coding agent is the LLM-driven environment you use to author runbooks: Claude Code in your terminal, Cursor in your editor, Microsoft Copilot Studio in your browser. The agent reads your codebase, calls tools, and writes code — and Ntropii is one of the tool surfaces it can call.

​
Why this matters

Writing a fund-ops runbook from scratch is slow. The author needs to know:
  • The exact warehouse layout (tables, columns, types) for the customer’s data platform
  • Which ntro SDK capabilities exist and how they compose
  • The customer’s chart of accounts, jurisdiction, and reporting obligations
  • The shape of incoming documents (invoice formats, bank statement layouts, property reports)
A coding agent with MCP + CLI access to Ntropii can answer most of those questions itself:
  • ntro://integrations/{id}/schemas lets it read the warehouse layout directly.
  • ntro://tenants/{slug} lets it read the tenant’s metadata and existing workflows.
  • ntro_workflow_create lets it deploy a generated runbook to Ntropii Tenant and bind it to an entity.
  • ntro workflow test (via the bash CLI) lets it run the runbook locally and iterate before deploying.
The result: a coding agent generates a working draft of a runbook in minutes, the human reviews and corrects it in a PR, the corrections feed the correction corpus, and future generations get better. This is the compiled-agent model in practice.

​
Two surfaces, both useful

MCP server

Structured tools, resources, and prompts. The agent calls ntro_workflow_create(...) and gets a typed response. Best for actions that need a contract.

CLI

The agent runs ntro workflow test in a bash shell and reads the output. Best for actions where the agent wants to see exactly what a human would see — progress bars, scenario summaries, error traces.
A well-wired coding agent has both. MCP for structured calls, CLI for terminal-shaped feedback loops.

​
Supported in this version

The MCP server is the same regardless of which client you use. The only difference is where the client expects to find its MCP config — ~/.claude/mcp.json for Claude Code, a Power Platform action for Copilot Studio, etc.

​
Authoring loop

Once your coding agent has MCP + CLI access, the typical loop looks like this:
1

Describe the workflow you want

“Generate a monthly NAV runbook for Acme SPV 1 — pull the trial balance from Snowflake, validate journal balance, and post the result back.” The agent reads ntro://tenants/acme-fund and ntro://integrations/{id}/schemas to ground itself.
2

The agent writes the runbook

Drafts runbooks/nav-monthly-acme/templates/workflow.py using the patterns in runbook-templates. References real ntro SDK capabilities — data.get_data_plane, accounting.validation, accounting.proposal — by reading the existing templates as examples.
3

You test locally

The agent runs ntro workflow test ./runbooks/nav-monthly-acme --scenario happy and reads the output. If a scenario fails, it iterates on the runbook.
4

You review and deploy

The agent opens a PR. You review the runbook, request changes if needed, merge. The agent (or you) runs ntro workflow create --path ./runbooks/<slug>/ --tenant <tenant> --entity <entity> — or asks the MCP ntro_workflow_create tool — to ship it.
The next two pages walk through the wiring for the two agents we currently document.