From idea to implementation, orchestrated.

Interactive TUI to orchestrate AI agents. Give it a PRD, get back architecture, test plan, security assessment, and working code.

pagent
 From idea to implementation, orchestrated.

 Input
 Select PRD or input file
 > 📁 examples/
   README.md
   examples/sample-prd.md

 Persona
   Minimal - MVP focus
 > Balanced - Standard
   Production - Enterprise

 Action
 > ▶ Run
   ⚙ Advanced...
   × Cancel

 ↑ up · ↓ down · / filter · enter select

Single-agent code generation has limits

When one agent handles everything - architecture, implementation, testing, security - quality suffers. Real teams don't work that way.

?

No architecture phase

Code gets written before API contracts, data models, or error handling strategies are defined. Inconsistent patterns across the codebase.

!

Shallow test coverage

Generated tests often cover happy paths only. Edge cases, error scenarios, and integration points get missed.

#

Security gaps

Input validation, rate limiting, and authorization checks are easy to overlook without a dedicated security review pass.

Structured handoffs between specialized agents

Each agent has a focused role and reads previous outputs as context. The architect defines the system before the implementer writes code.

~
PRD.md
Your requirements
~
Specification Phase
architect
qa
security
~
Implementation Phase
implementer
verifier
~
~
outputs/
Specs + Code

Five specialized roles

Each agent receives a focused prompt and produces a specific deliverable. Downstream agents read upstream outputs as context.

~

architect

architecture.md

REST/GraphQL endpoints, DB schema, error handling patterns, deployment config

~

qa

test-plan.md

Test cases per endpoint, edge cases to cover, integration test scenarios

!

security

security-assessment.md

OWASP checks, auth flow review, input validation requirements, rate limit specs

#

implementer

code/*

Reads all specs, writes handlers, services, repos, middleware, Dockerfile

~

verifier

*_test.go

Writes tests from test-plan.md, checks security requirements, validates build

Practical workflow automation

~

Multi-LLM support

Currently supports Claude Code, with Gemini CLI and Codex CLI coming soon. Uses your existing authentication - no additional API keys required.

||

Parallel execution

QA and security agents run concurrently after architecture completes. Dependency-aware scheduling minimizes total runtime.

~

Incremental regeneration

Content-hash based change detection. Only affected outputs get regenerated when inputs change.

Interactive TUI

The primary interface. Auto-discovers PRD files, lets you pick personas and agents visually, with keyboard navigation and file browser.

~

Modify existing codebases

Point pagent at an existing project and it will respect your patterns, conventions, and directory structure.

~

Configurable tech stack

Define your preferences once in .pagent/config.yaml. Language, API style, cloud provider, database, CI/CD tooling.

MCP Server

Run as an MCP server for Claude Desktop integration. Supports stdio and HTTP transports with optional OAuth.

Match rigor to your project needs

Different projects need different levels of thoroughness. Choose the right trade-off between speed and completeness.

--persona minimal

Minimal

POC / hackathon

  • Happy path handlers only
  • Basic input validation
  • One test per endpoint
  • No observability
--persona production

Production

Enterprise / regulated

  • STRIDE threat modeling
  • 95% coverage on critical paths
  • Audit logs, rate limiting
  • OpenTelemetry traces

Try it out

1 Install pagent
brew install tuannvm/mcp/pagent

Alternative: download binary or build from source

2 Launch the TUI
pagent ui

Or use CLI: pagent run ./prd.md -s -v

3 Check your outputs
ls outputs/
# architecture.md test-plan.md security-assessment.md
# code/ directory generated when implementer agent runs