A2Flow

A2Flow is an open-source, chat-driven workflow engine from the GitHub user kaitoy, built around an AI agent. It targets ITSM-style procedures such as service requests, change execution, and announcements. The agent reads a procedure, plans it as a task graph, pauses for human approval, and executes the remaining steps.

What It Is

A2Flow is a workflow engine whose runtime is an agent rather than a fixed flowchart. Procedures are written as Agent Skills, which are SKILL.md files in a Git repository. In a design session the agent turns a Skill and a user's prompt into a task graph with explicit dependencies. Once a human publishes that graph, it is kept as a workflow, and every later run follows the same steps.

How a Run Works

Each run has three phases: plan, approve, execute. Nothing runs until a human publishes the design. Tool calls that change something require the designated approver, who sees Approve / Reject / Return controls in the chat. During execution the agent walks the graph in dependency order and calls only the MCP tools bound to the current task. It asks questions through forms rendered in the chat via the A2UI protocol. The applicant, the approver, and the agent share one conversation thread.

Security and Governance

The README describes per-task short-lived X.509 certificates over the granted tools, enforced by an MCP gateway. Third-party MCP server code runs in a separate proxy, with each launch in its own bubblewrap sandbox. Other documented features include tool mocks for side-effect-free testing, named secrets (database or HashiCorp Vault), append-only audit logs, multi-tenant role-based authorization, notifications, and an admin UI.

Setup Path

The stack (PostgreSQL, backend, email worker, MCP proxy, frontend) starts with Docker Compose using a Google API key. Gemini is the default model, and OpenAI and Anthropic models are supported via LiteLLM. The backend is FastAPI with Google ADK, and the frontend is a Next.js chat UI. It is released under Apache License 2.0.