BradGroux/veritas-kanban

What it solves

Veritas Kanban is a local-first task management system designed to bridge the gap between human planning and AI agent execution. It prevents the chaos of unsupervised AI agents by providing a structured Kanban board that acts as the source of truth, combined with governance tools to monitor, restrict, and approve agent actions.

How it works

The system uses a local-first approach where tasks are stored as Markdown files and settings as JSON. It provides a visual Kanban interface for humans to organize work, which can then be connected to various AI agent runners (such as OpenAI Codex or HermesAgent). It employs a "layered" architecture: users start with a simple board and can optionally add a CLI, an MCP server for tool access, or complex workflow engines for multi-agent orchestration.

Who it’s for

It is built for developers and teams who want to use autonomous AI coding agents but require strict control, audit trails, and human-in-the-loop approvals to ensure safety and code quality.

Highlights

• Agent Governance: Includes a policy engine for allowing/denying specific tools, sandbox presets for filesystem/network access, and behavioral drift detection.
• Multi-Agent Orchestration: Supports spawning multiple agents per task with different roles (Intern, Specialist, Lead) and a shared "Squad Chat" for coordination.
• Agentic Workflows: A YAML-based workflow engine that allows for sequential or parallel agent pipelines with human approval gates.
• Git-Native Integration: Uses isolated Git worktrees per task to prevent branch conflicts and integrates directly with GitHub PRs and Issues.
• MCP Server: Provides 42 tools via the Model Context Protocol to allow agents to interact with the board and system.
• Local-First Storage: Defaults to Markdown and JSON files, making it easy to back up and independent of heavy infrastructure.