Turn AI-generated code into shipped products
Agent-Dev manages the full lifecycle from requirement to production. Four real projects have been delivered end-to-end on real cloud infrastructure, with code, cloud resources, and domains fully owned by you.
Golden Path
The complete pipeline from requirement to delivery report
- Requirement
Describe goal and scope in natural language
- Clarification
Ask only about real uncertain decisions
- Blueprint
Generate structured product spec and acceptance criteria
- Resource ownership
Confirm GitHub/cloud ownership with named approval
- Local Apply
Generate engineering baseline and Git init in isolated workspace
- Quality gate
lint / typecheck / unit / build actually executed
- Feature Task
Define feature goal and boundary; execute after human approval
- Agent implements
Replaceable runtime (Codex/OpenCode) codes in sandbox
- PR + CI
Platform pushes and opens PR; GitHub Actions quality gate
- Dual Preview
Vercel API + Cloudflare Pages joint deployment
- Production release
Checkout from production branch; release after two human gates
- Delivery report
Evidence chain, observations, and residual risks
Why Agent-Dev
Not a faster code generator, but accountability for the full delivery
User-owned
Code, repos, data, domains, cloud resources, and Blueprint belong to the user. Four real projects have verified that products remain independently buildable, deployable, and maintainable after leaving Agent-Dev.
Six product templates
Web SaaS, landing pages, browser extensions (MV3), desktop (Tauri/Electron), mobile (Expo), and MCP Server. Each artifact passes both local and real GitHub Actions quality gates.
Real gates
An LLM may propose, but cannot obtain production permission via prompt. Production release requires two human gates (request + named approval), published from a production branch checkout.
Evidence-based delivery
Delivery status is grounded in GitHub Checks, Deployment Records, Preview URLs, smoke tests, and real observations, not the Agent's natural-language summary. Unexecuted verifications cannot be marked as passed.
Compared to adjacent products
Agent-Dev doesn't compete on who generates pages faster
| Dimension | Agent-Dev | Replit Agent | Lovable | GitHub Spec Kit |
|---|---|---|---|---|
| Unit managed | Product lifecycle | Single session | Single generation | Specs and tasks |
| User-owned resources | Yes, own GitHub/cloud | No, Replit-hosted | Partial | Yes |
| Replaceable Agent | Yes, 6 runtimes | No | No | Partial |
| Product types | Six real templates | Web app | Web app | N/A |
| Portable providers | Yes, controlled migration | No | No | N/A |
| Production permission | Policy + two human gates | Platform-managed | Platform-managed | Manual gate |
| Delivery evidence | External fact chain + observations | Platform logs | Preview | Task status |
| Long-term maintenance | Evidence retained + artifact drift check | Limited | Limited | N/A |
Responsibility Boundary
Transparent split between user and platform builds trust
User is responsible for
- →Who the product serves and what problem it solves
- →Feature scope, experience, aesthetics, and business priority
- →Cost, privacy, data, and external-impact decisions
- →Preview experience acceptance
- →Production release and final Approval
Agent-Dev is responsible for
- →Provide six verified product Blueprints and default standards
- →Turn requirements into structured Blueprints, acceptance criteria, and Feature Tasks
- →Call replaceable Agent Runtimes (Codex/OpenCode/Claude, etc.) to implement, test, and fix code
- →Orchestrate real GitHub, Vercel, and Cloudflare providers; Supabase is a guided manual setup
- →Manage credentials and env vars; secrets are never returned via API or written to DB
- →Enforce deterministic Policy to constrain Agent permissions and production access
- →Orchestrate Dual Preview deployment and joint smoke tests
- →Aggregate real CI, deploy, test, and manual-acceptance evidence, recording observations
- →Validate workspace artifact drift, detect stale config and maintenance tasks
- →Report clear delivery status and residual risks
Agent-Dev guarantees process and evidence integrity, not market success.
Product Constitution Excerpts
Core principles that constrain platform behavior
Defaults over choices
Beginner mode decides ~90% of engineering questions for the user, asking only about product, cost, privacy, or risk. Options are escape hatches, not the default experience.
AI reasons, Policy authorizes
An LLM may propose, but cannot obtain production permission via prompt. Deterministic Policy and provider permissions are the final constraints.
External facts over Agent self-report
Delivery status is grounded in GitHub Checks, Deployment Records, Preview URLs, and smoke tests, not the Agent's natural-language summary.
User owns the product
Code, repos, data, domains, cloud resources, and Blueprint belong to the user. Models and providers are replaceable.
Frequently Asked Questions
Key questions about ownership, security, and automation
What is Agent-Dev?
Agent-Dev is an Agentic Product Delivery Platform. It is not a replacement for Codex or Claude Code, nor a source-code-only app builder. It manages the full product lifecycle from requirement to production.
Who owns my code and cloud resources?
Code, repos, data, domains, cloud resources, and Blueprint all belong to the user. After leaving Agent-Dev, the product still builds, deploys, and maintains independently. Models and providers are replaceable.
How does Agent-Dev control AI production permissions?
An LLM may propose, but cannot obtain production permission via prompt. Deterministic Policy, GitHub Rulesets, Environment Approval, and provider permissions are the final constraints. Production release requires two human gates (request + named approval), published from a checkout of the recorded repository's production branch. Autonomy scales per project and per action; there is no single risk-blind "full auto" switch.
How is delivery actually verified as complete?
Delivery status is grounded in GitHub Checks, Deployment Records, Preview URLs, database state, smoke tests, and manual acceptance, not the Agent's natural-language summary. Evidence records real observations (HTTP status, content-type, measured CORS headers), not judgment constants like "passed". Unexecuted verifications cannot be marked as passed.
How is Agent-Dev different from Replit Agent or Lovable?
Replit Agent and Lovable focus on conversational generation and one-click hosting on platform-owned resources. Agent-Dev's minimum unit is the product lifecycle: it uses user-owned GitHub/cloud resources, with replaceable Agents and providers, and is accountable for full delivery and long-term maintenance.
What tech stack does the first version support?
The Web SaaS Golden Path is fixed: React + Vite + TypeScript frontend, Hono API, Cloudflare Pages for frontend hosting, Vercel Functions for API hosting, and GitHub Actions for CI. Supabase provides DB and auth through a guided manual setup. The Agent Runtime runs on your own machine, selectable among installed coding agents such as Codex, OpenCode, Claude Code, and Aider.
Which product types are supported?
All six product types generate real, buildable project templates: Web SaaS, landing pages, browser extensions (MV3), desktop apps (Tauri v2 / Electron dual shell), mobile (Expo SDK 52), and MCP Server. Each artifact passes its local quality gate, and the generated GitHub Actions workflows have been verified green on real CI. Web SaaS follows the full cloud delivery pipeline; other types ship locally buildable artifacts by design, with signing, notarization, and store submission as manual steps.
Are there real delivery examples?
Four real projects have been delivered end-to-end on real cloud infrastructure: Receipt Test (receipt management page), Workspace Verify Fresh (API version endpoint), Link Vault (link-saving API + page), and MCP Word Tools (MCP server, first non-web-saas delivery). Each project completed the full Blueprint → Preview → Production cycle, with PRs merged after GitHub Actions quality gates, and production APIs and pages publicly accessible and independently verified outside the platform report. The four validation rounds fixed 29 real defects.
How does Agent-Dev run today?
v0.2 is a local-first application, not a hosted SaaS: the control plane runs on your own machine and reuses your existing GitHub, Vercel, and Cloudflare CLI logins plus a local coding agent. It is still a v0.2 experiment; the core delivery pipeline has been verified on real cloud infrastructure, but it is not yet released as a production-ready stable version.
Start your first product delivery
Get a product baseline you own and can keep developing