OpenJCockpit is an open-source AI Software Factory: an AI Orchestration Dashboard and backend platform for governed agentic software development. It coordinates AI agents in structured, specification-driven workflows, with policy checks, human approval gates and reviewable Git delivery around them.
This page explains the concepts, separates the long-term vision from what is implemented today, and points to the detailed documentation.
A single coding agent can generate code. Delivering production-quality software takes more: requirements, architecture, coordinated implementation, testing, review, policy and controlled delivery. OpenJCockpit’s approach is to put those activities into orchestrated workflows of specialised agents, each with explicit permissions, and to keep humans in control of the decisions that matter.
| From | Towards | How OpenJCockpit approaches it today |
|---|---|---|
| AI-assisted coding | Orchestrated AI software engineering | Workflows chain role-based agents (analyst, architect, developer, QA), each building on the previous outputs. |
| Individual AI agents | Coordinated development workflows | A central agent catalogue, workflow groups, workflow chaining and a per-project spec queue. |
| Unstructured prompting | Specification-driven engineering | Specs in the repository are the input contract for implementation workflows. |
| Autonomous code generation | Governed, reviewable delivery | OPA policy checks, tool allowlists, guardrails, approval gates and one Git branch per run with checkpoint commits. |
| Experimental AI workflows | Enterprise engineering practices | Keycloak identity, audit trails, CI with a 100% backend line-coverage gate, documented limitations. Production hardening is still ahead. |
| Concept | What it is | Read more |
|---|---|---|
| AI Orchestration Dashboard | The web UI where you select a project, design and start workflows, follow agent runs, take approval decisions and manage specs, skills and queues. | Below |
| Project | A software project with its Git repository and credentials. Runs on a project work in that repository. | Git integration |
| Spec | A Markdown specification with requirements and acceptance criteria, stored in the project’s specs/ folder. |
Spec-driven development |
| Agent | A role-based AI agent (for example Software Developer) with a model alias, instructions, skills, tools and guardrails. Defined locally in Markdown or served over A2A. | Agents |
| Workflow | An ordered chain of agents, optionally with approval gates and workflow orbs. | Agents and workflows |
| Approval gate | A human decision point inside a run: accept, deny or accept with comments. | Approval gates |
| Policy | OPA rules evaluated for workflow starts, agents and every tool call. | Policy enforcement |
| Tool | A capability an agent may call, such as reading or writing repository files, via MCP. | MCP tools |
| Skill | A reusable instruction package for agents (Agent Skills format). | Skills Hub |
| Spec queue | A per-project queue that runs specs one at a time, optionally auto-merging results. | Spec queue |
The dashboard is a React application behind Keycloak login. After signing in, you see an animated landing hub (an intro greeting with your first name, a top bar with Profile, Notifications, Messages and Logout, and two portals: Business and Implementation). From there you work with:
| Area | What you can do |
|---|---|
| Project selection and settings | Choose a project; manage its repository and Git credentials. |
| Workflow dashboard and overview | See the workflows available for the project and start them, with a prompt when the workflow asks for one. |
| Workflow Design | Compose workflows from catalogue agents (Local / LiteLLM sections), set the execution order, add approval gates and workflow orbs, group workflows, export/import JSON. Browse agents, skills and A2A services. |
| Workflow Configuration | Configure the OPA policy client and A2A services. |
| Workflow execution and history | Follow a run live: agent sequence, agent logs and outputs, events, policy decisions, Git branch and status (“run state lost” for unknown runs). |
| Approval decisions | Review the approval context (including changed files) and accept, deny or accept with comments. |
| Spec files | Browse the project’s specs, run the spec-init flow, create a spec from a prompt. |
| Spec queue | Enqueue, reorder, pause, retry or skip specs; admins control auto-merge. |
| Skills Hub | Browse skills and register Skills Marketplace connections. |
The UI uses a self-authored, futuristic animated/framed look, also applied to the custom Keycloak login
theme. See frontend/libs/ui/README.md.
The original README listed the following under “What is in this project?”. It is preserved here with corrected paths:
frontend/libs/ui/README.md).keycloak-js.Hi <name>!, where the name is retrieved from the Keycloak token (the code uses
the first name).infrastructure/keycloak/openjcockpit-realm.json.infrastructure/themes/openjcockpit/login.openjcockpit uses loginTheme: openjcockpit by default.OpenJCockpit is ambitious in direction and explicit about its current state.
| Capability | Status |
|---|---|
| Orchestrated, multi-agent workflows from a dashboard | Implemented |
| Agent catalogue with local Markdown agents and A2A agents via LiteLLM | Implemented |
| Spec-driven workflows (spec init, spec from prompt, implement, realise) | Implemented |
| OPA policy checks for workflow starts, agents and tool calls; audit trail | Implemented; default policy is a permissive starting point |
| Human approval gates with bounded feedback loops | Implemented; paused runs are in memory only |
| Git delivery: one branch per run with checkpoint commits, server-side credentials | Implemented |
| Automatic pull request creation for workflow runs | Not in the default workflows (a PR tool exists in the Git MCP server) |
| Spec queue with optional auto-merge | Implemented; auto-merge needs a run-reported PR |
| Skills Marketplace integration | Early: placeholder contract, dev-only mock |
| Fine-grained roles, multi-tenancy, production deployment | Not yet |
| Java application development | Primary focus |
| Other programming languages | Extensibility opportunity: agents work on repository files and are configurable, but non-Java projects are not a verified target yet |
OpenJCockpit is built in Java (Spring Boot, Embabel, Spring AI) and focuses primarily on Java application development. The architecture does not hard-wire a language into the workflow engine: agents are configurable Markdown definitions with instructions, skills and tools, and the Git tools operate on plain repository files. That makes support for other languages a natural extension point, through new agent definitions, skills and workflows, rather than a rewrite. It is not yet a proven capability.