Direct answers.

Specship is a spec-driven AI engineer that picks up tickets from the PM tool your team already uses (GitHub Issues, ClickUp, Jira, Linear, or Specship's own kanban), writes failing tests first, writes the code to make them pass, refactors, and opens a pull request. Test-driven development is enforced — it is not a configurable toggle.

Cursor and Aider are pair-programmers in an editor — a human drives. Devin attempts to be an end-to-end engineer in its own VM. Specship sits in the middle: it lives inside your existing PM + Git workflow, follows strict TDD discipline, and is opt-in per ticket via assignment or label. The agent never works on a ticket the human didn't explicitly hand off.

Two layers. The orchestrator (queueing, state, dashboard, webhooks, integrations, and MCP adapters) runs in our cloud. Coding jobs run against isolated working trees and can use supported engines such as Claude Sonnet 4.6 and OpenAI Codex GPT-5.5 where enabled. Billing and execution controls are still being tuned during private beta, with budget caps and pause controls designed to reduce surprise usage.

Hard failures are first-class. After three failed self-correction attempts on the green TDD phase, or if the same file gets rewritten more than 5 times (loop detection), the agent stops, posts a structured comment listing what it tried and what the human needs to do, unassigns itself, and labels the ticket "shipd: blocked". The branch is left intact for human review.

Yes — the agent has to read and write your repo. Live integrations use scoped OAuth and native adapters; ClickUp tokens are encrypted in Specship’s vault and decrypted only for scoped integration calls. The bot account commits only to branches prefixed shipd/, and your default branch is policy-protected from agent writes.

A "ticket execution" is counted only when the agent picks up a ticket AND opens at least one commit. Clarification rounds (questions, no code yet) and post-PR review iterations are always free. Tier prices and execution caps are still being tuned during the private beta — join the waitlist to lock in the launch rate. Overages will be opt-in with a monthly cap you set.

The agent inherits whatever your project uses. Verified end-to-end: TypeScript/JavaScript with Jest or Vitest. Strong support: Python with Pytest, Ruby with RSpec, Go with the stdlib testing package. The framework is detected from the project's manifest (package.json, requirements.txt, etc.) during the Project Summary Card generation step.

Private-beta teams can choose supported coding engines where enabled, including Claude Sonnet 4.6 and OpenAI Codex GPT-5.5. Bring-your-own-key and metered billing options depend on the workspace setup during beta.

Set the project's reviewerRequirement to maintainer_approval. Every new GitHub Issue is scored by Haiku ("AI-fixable right now?") with a score 0–1 and a verdict (fixable / risky / human_only). High-scoring issues land in your Triage Queue. You explicitly approve which ones the agent attempts — the agent never acts on an unapproved issue. PRs are opened as drafts; the agent never marks them ready-for-review.

Next.js 16 (App Router) for the dashboard + API. PostgreSQL via Prisma. Redis + BullMQ for the job queue. Native integrations and MCP adapters for third-party I/O. Supported coding engines include Claude Sonnet 4.6 and OpenAI Codex GPT-5.5 where enabled. Jobs run real project test suites, can provision ephemeral Postgres for migrations, rebase onto main before PR, and stream activity to the dashboard via SSE backed by Redis pub/sub.