Skip to main content
Human guide: work with your agent →
Install Deepline in the agent runtime and produce one reviewed pilot in the intended workspace.
Use this when you want an agent to run GTM work without opening the Deepline UI.

​
Pick the integration path

The rule is simple: if the agent can run shell commands, it can run Deepline. Use skills where the agent supports them. Use the plain CLI everywhere else.

​
Ask the agent to install Deepline

Give the agent this instruction:
The guide makes the agent check its environment, install deepline@latest, run deepline setup, handle the persistent project-local fallback in a restricted sandbox, complete browser authorization, and verify CLI, skills, auth, and workspace state. For a direct terminal install: check npm --version and use the matching command. For npm 11.16 or newer:
For older npm:
In a sandbox with no writable global npm prefix, install into the project instead:
deepline setup installs the hosted skill bundle for the detected agent and keeps it current. The hosted installer supports Codex, Claude Code, Cursor, Gemini CLI, and Antigravity. For an unknown agent it installs the bundle for all supported targets. Hermes remains CLI-only. OpenClaw uses the CLI or MCP and does not have a hosted Deepline skill target.

​
Give the agent a concrete job

Create leads-pilot.csv with CSV-safe record selection before the pilot commands. Good agent instructions name the file, output file, run name, row range, and tool payload.
If the pilot looks right, run the same checked Play on the full file:
Do not ask the agent to guess flags or Play names. Have it inspect the Play, prepare the pilot, and run the exact sequence above.

​
Review the pilot before expanding it

Ask the agent to report the following before processing the full input:
  • The exact Play or tool contract and the identifiers used to match records.
  • Accepted, unresolved, and failed rows, with evidence for the acceptance rule.
  • Recorded Deepline usage and the selected managed or BYOK mode.
  • Proposed external changes and the fields the workflow is permitted to write.
  • The retry and review policy for uncertain destination outcomes.
Include representative missing, ambiguous, and conflicting inputs in the pilot. A successful authoring check does not prove the business result. Keep external writes disabled in the authored logic until their policy is ready. This is a workflow control you implement, not a universal Play dry-run flag. Save the run ID so later inspection observes the same work.

​
Use the GTM agent repo

For repo-backed agent workflows, start from the Deepline GTM agent project:

deepline-gtm-agent

Reference repo for improving Deepline GTM agent workflows across coding agents.
Use that repo when you want versioned prompts, reusable workflows, or a shared agent setup that works across Codex, Claude Code, and other shell-capable coding agents.

​
Troubleshooting

Ask the agent to follow https://code.deepline.com/INSTALL.md in that environment. It will use a persistent project-local install when sandbox policy blocks a global installation.
Read the failed phase and exact retry command from deepline setup --json. If setup reports a diagnostic command, run it. Otherwise, in a project-local installation, use ./.deepline/runtime/bin/deepline doctor --scope local --json. after a global install, use deepline doctor --json.
Stop and check the internal content canonicals before running it. For enrichment, use the command shape on this page.

​
Verify this task

Confirm CLI availability and effective authorization. Review pilot source IDs, accepted outcomes, exceptions, and scoped usage before scaling. Ask your agent to explain a failed check before repeating work or expanding the scope.