Ship

Ship is Teploy's autonomous coding agent, self-hosted: issues in, pull requests out, on your own servers.

The agent acts by writing code, not by emitting structured tool-call JSON. Each turn it thinks, then runs one fenced code block (bash or Python) in a sandbox, observes the real output, and iterates until the task is done. This is the CodeAct strategy, implemented against Teploy's own stack rather than ported.

The guarantees

• A pull request that carries its evidence. With SHIP_TESTS=1 the test suite runs in the workspace after the agent stops and before the push — "tests passed" describes the code that became the PR, not the agent's account of it. A failing suite still publishes the PR, marked, with its output.
• Caps and approvals. Spend limits, plan preview and review gates. Destructive, network and privilege actions park the run for a human decision; an agent that could approve its own drafts would be no gate.
• Durable runs. A run is a workflow over a file-backed event log. A crashed run replays completed turns and continues — no re-calling the model, no re-running commands. A parked run holds zero processes.
• Model-agnostic. Two adapter shapes (Anthropic-compatible and OpenAI-compatible), direct or through a gateway you run, with your credentials and no markup. Validated and prompt-tuned on GLM; measured results per model are published in the repository's docs/MODELS.md.

Where to go next

• Quick start — first run, live and durable
• Harnesses — native loop, Claude Code, OpenCode
• Self-hosting — install, workers, webhooks, preview deploys

The repository is github.com/useteploy/teploy-ship.