POCLI logo

Define the intent. Direct the agents. Deliver verified software.

An open-source, Linux-native terminal environment (TUI & CLI) where you act as the Product Owner at the helm of a governed multi-agent engineering team.

From Theory to Real Practice: Learning AI Engineering

Everything documented across AI Engineering — from Harness Engineering, Agent Loops, and Spec-Driven Development to Model Routing, AI Infrastructure, and AI Product Governance — is actively materializing inside POCLI.

Rather than keeping these concepts as isolated notes or toy scripts, POCLI serves as a real, modern engineering codebase and hands-on learning platform where every architectural pattern of modern AI systems is implemented end-to-end:

  • Application Development: Spec-driven workflows with traceable normative bundles (REQ, ADR, ARCH, SPEC) and Product Owner human-in-the-loop gates.
  • Model Development & Routing: Multi-provider gateway (OpenRouter, Groq, Cerebras, Gemini, Ollama, OpenCode Zen, Anthropic, OpenAI) with sliding-window quotas, Bandit presets, and automatic failover.
  • AI Infrastructure: Minimal token-efficient harness, SQLite WAL persistence, 3-tier context window assembly, 4-layer memory, lossless compaction (/compact), and local Hybrid RAG (SHA-256 + FTS5 + vector similarity).
  • AI Governance: Deterministic 8-phase SDLC state machine (AuthorityGuard), single-writer database coordinator (PoCliCoordinator), Linux Bubblewrap (bwrap) / Podman OCI process isolation, and independent verification gates.

Terminal Command Center in Action

The walkthrough below runs the real Ink TUI (CommandCenterApp): creating a project (/project), assigning models to all agents (/models all), collaborating with the Software Architect, running autonomous ARQ → DEV → QA → ARQ execution and verification, cycling agent channels (Ctrl+N), inspecting the 4-panel Multi-View Command Center (Tab), and generating an executive HTML status report (/agents pm).

POCLI Terminal Command Center Demo

The Virtual Engineering Team

Each specialist runs inside an isolated Linux OS worker process over JSON-RPC 2.0 stdio with strict role-bounded tools (allowedTools) and zero cross-conversation context bloat:

Role Responsibility Key Outputs & Verification
Product Owner (You) Defines product intent, reviews specs, and approves gates /project, /models, /approve, /accept
Software Architect Translates intent into requirements, ADRs, architecture, and technical specs REQ-*, ADR-*, ARCH-001, SPEC-001, [ARQ → DEV]
Senior Developer Implements software, writes unit tests and Gherkin contracts inside workspace repos/, CODE-001, DEV-GHERKIN-001.feature, [DEV → QA]
Independent QA Engineer Provisions isolated .venv, executes Robot Framework & Mutation Testing QA-001, TC-MUT-001, bidirectional [QA → DEV] loop
Project Manager (PM) Queries across roles and compiles full executive HTML delivery reports PM-001, deliveries/project-status-report.html

Key Engineering Highlights

Capability How POCLI Implements It Reference on This Site
Minimal Harness & Loop JSON-RPC worker loop with agent.streamChunk TUI streaming and MAX_QA_CORRECTION_CYCLES = 3 Harness Engineering · Agent Loop
Context & Memory Engine 3-Tier Context Window + 4-Layer Memory + lossless /compact checkpoints in SQLite (context_compactions) AI Agent
Local Hybrid RAG NotebookRagEngine with incremental SHA-256 change tracking, SQLite FTS5, and vector retrieval AI Infrastructure
Governance & Isolation AuthorityGuard SDLC state enforcement, PoCliCoordinator sole write owner, bwrap / podman sandboxing AI Governance
Headless E2E & Standards Built-in McpRegistry (JSON-RPC stdio MCP client), pocli db SQL REPL, pocli e2e, and pocli tui-test Tools · OpenCode

Quick Start (Linux)

Install POCLI in one command (Node.js >= 22.5 LTS):

curl -fsSL https://artificial-intelligence-engineering.github.io/pocli/install.sh | sh

To explore the complete documentation, interactive AI Agent Harness Reference Matrix, and bilingual guide: