Skip to main content
Deepline’s CLI emits inspectable output, composes with shell tools, and fails loudly because those are interfaces a coding agent can act on.

​
The operating principles

​
1. Discover the live contract

Agents should not guess a tool identifier, Play input, provider route, or price from memory. They search and describe the current catalog first:
Use --json when the next step is programmatic selection. Human-readable output remains useful for direct review.

​
2. Keep multi-step work in a Play

A shell command starts a Play run. The command invokes a maintained or owned Play. The Play defines typed inputs, outputs, review boundaries, triggers, and versioned source.
Inspect the tool contract before a one-time deepline tools execute operation. Use a Play for work that has multiple steps, validation, review, or external writes.

​
3. Pilot the production-shaped workflow

Create a small input that represents the production data. Run the production Play with this pilot input:
Review output fields, provenance, usage, errors, and downstream effects before running the full input.

​
4. Make failure and state visible

Agent-callable interfaces expose Play run status, output, usage, and errors. Input, contract, provider, and billing errors must not change the Play without notice.

​
5. Preserve operator choice

Use the CLI with coding agents that can run shell commands. Use the SDK to author typed Plays or call Deepline from an application. Use the API or MCP when the runtime requires that interface.

​
What the agent should review

Before a full run or external write, confirm:
  • the exact Play or tool contract
  • representative pilot results
  • provider provenance and validation
  • BYOK or managed billing behavior
  • no-result and failure states
  • idempotency and review boundaries for CRM or sequencer writes
Use a workflow that the agent can inspect, run, and explain.

​
Further reading

  • MCP setup — connect ChatGPT Work, Claude Desktop, or another MCP client
  • Quickstart — install Deepline and run a maintained Play
  • Coding Agents — install Deepline for shell-capable coding agents
  • Claude Code skills — teach Claude Code about Deepline workflows