Harness · Claude Code

Claude Code on the Software Lifecycle Cloud

Claude Code on KERNSTACK means Anthropic’s terminal coding agent working in your project directory: reading the codebase, editing stack.json next to the code it describes, and driving the kernstack CLI itself. Because plan output is available as JSON, the agent can check its own work against the platform’s diff before a person ever sees it — and the deploy gate decides what happens next.

A terminal agent meets a declarative target

Claude Code’s strength is sustained, multi-file work from the command line: refactors, migrations, test repair. KERNSTACK gives that work an unusually checkable target. The agent edits a function, declares its route in stack.json, runs kernstack plan --json, and reads a structured answer — these resources will be created, these updated, these deleted — instead of inferring the blast radius from code alone.

That loop is the difference between an agent that writes code and an agent that operates a system. The plan is the same artifact a human reviewer approves, so nothing the agent concluded privately is lost: the diff it read is the diff you review.

Where the human checkpoint lives

Nothing about a terminal agent should end in an unattended production change, and on KERNSTACK it does not have to. Projects can require review-approved source before deploy (min_source_assurance), destructive operations demand an explicit confirmation flag, and work done in a sandbox stays on its agent/ branch with its own agent- environment until a person promotes it.

  • kernstack plan --json: the agent verifies its change set before presenting it
  • kernstack git checkout agent/: a person pulls the agent’s exact branch and runs it locally
  • kernstack deploy --confirm-destructive: deletes never ride along silently in an agent’s plan

Supported Models & Workflow

Supported Models & Engines: Claude (Opus, Sonnet, Haiku — your choice in Claude Code)

Frequently asked questions

Technically the CLI is right there in its terminal — which is why the platform, not the prompt, carries the boundary. Set the deploy gate to require review-approved source and the agent’s deploys stop at the gate until a person approves. Sandboxed work additionally stays on its agent branch and environment.

Agents reason better over structured facts than terminal prose. kernstack plan --json returns the change set — creates, updates, deletes — in a form Claude Code can assert against your instructions (“no deletes”, “only this function”) and report back accurately, instead of paraphrasing scrollback.

It needs the repository locally and the kernstack CLI; hosted git is what adds the agent-branch workflow — sandboxed branches, agent-<id> demo environments, and clean checkout of agent work. A plain local repo can still plan and deploy.

Ship your stack on the Software Lifecycle Cloud.

Closed preview on dedicated European infrastructure. Bring a repository or start from a blueprint.