AI Agent Overview

DevFlow is built for AI-assisted development. It connects to AI coding agents through a standard protocol, giving the agent structured access to your flows, tasks, and project context. The agent works within the same workflow you see in the UI — planning, implementing, and reviewing — while you stay in control at every step.

How It Works

DevFlow communicates with AI agents through the Model Context Protocol (MCP). This gives the agent a set of tools to interact with your project:

• Read flows and tasks — The agent sees what needs to be done
• Create plans and tasks — The agent breaks work into structured steps
• Update progress — The agent reports what it has done
• Submit for review — The agent hands work back to you for approval

The agent never works in a vacuum. Every action it takes is visible in DevFlow, and every major decision requires your approval.

Supported AI Clients

DevFlow works with any MCP-compatible AI coding assistant:

Client	Status
Claude Code	Fully supported
Cursor	Supported
Gemini CLI	Supported
Windsurf	Supported

See MCP Setup for installation instructions per client.

Pipeline Rules

The pipeline controls how agent and human collaborate at each step. Every step has four independent settings:

Setting	Description	Values
Actor	Who does the work?	human, agent, both, auto
Transition Policy	Who can advance to the next step?	human_only, agent_only, human_or_agent, auto
Kind	What type of step is it?	work, review, handoff, terminal
Skippable	Can this step be skipped?	true, false

This separation means you can configure steps where the human does the work but the agent is allowed to move forward (actor: human, transitionPolicy: human_or_agent), or where the agent works but a human must approve before continuing (actor: agent, transitionPolicy: human_only).

Presets

Three built-in presets configure the pipeline for common scenarios:

Preset	Description
Off	Minimal process — agent runs freely, only execution step active
Balanced	Standard workflow — agent plans, human approves, agent implements
Strict	Maximum control — reviews at every stage, all steps active

You can customize individual steps after applying a preset.

Agent Status Tracking

You can see what the agent is doing at any time. The agent reports its current status as it works:

Status	Meaning
Idle	Not currently working on anything
Analyzing	Reading requirements and exploring the codebase
Planning	Creating an implementation plan
Implementing	Writing code
Testing	Running and verifying tests
Reviewing	Checking its own work before submitting

Status Indicator

The header bar shows an agent status indicator whenever the agent is actively working. Click it to open a dropdown with details about:

• Which flow the agent is working on
• What the agent is currently doing
• How long the current session has been running

This gives you a quick glance at agent activity without leaving your current view.

Pending Questions

Sometimes the agent needs clarification before it can continue. When this happens, you receive a notification in DevFlow. The agent pauses and waits for your response before proceeding.

This keeps the human-in-the-loop principle intact — the agent does not guess when it is unsure.

Agent Sessions

Every unit of work the agent performs is logged as an agent session. A session records:

• Start and end time — When the agent began and finished working
• Activity log — Key decisions, progress updates, and any issues encountered
• Summary — A brief description of what was accomplished

You can view all agent sessions for a flow in the flow detail view. This gives you a complete history of the AI’s work, which is useful for auditing, debugging, or understanding how a feature was built.

Agent Time Tracking

AI work time is tracked separately from your own time. This distinction matters because:

• You can see how much time the AI spent versus how much you spent
• Time reports clearly separate human and machine effort
• You get an accurate picture of total project effort

Agent time is calculated automatically from session start and end timestamps. It appears alongside your manual time entries in the flow detail and project dashboard.

Two Modes: Interactive and Runner

DevFlow supports two ways of working with AI:

Interactive Mode (MCP)

You work in your IDE (Claude Code, Cursor, Codex, Gemini) and the AI follows DevFlow’s pipeline step by step. You are actively involved — the MCP server tracks progress and enforces rules.

Runner Mode (Autonomous)

The DevFlow Runner is a local CLI tool that autonomously executes flows. It spawns AI tools, objectively verifies results (exit codes, not AI opinion), and self-repairs on failure.

Install

Install via the Agent Hub in DevFlow (Settings → Agent Hub → Install Runner). The wizard generates a setup command:

npx https://api.app.dev-flow.tech/api/downloads/devflow-runner-latest.tgz setup --token dfr_xxx --url https://api.app.dev-flow.tech

Usage

# Start watch mode — connects via Socket.IO + polls for new work
devflow-runner watch

# Run a specific flow
devflow-runner run df-42

# Run all ready flows
devflow-runner run --all

# Run until next human gate
devflow-runner run df-42 --until-gate

# Dry run — see what would happen
devflow-runner run df-42 --dry-run

Configuration

The Runner is configured automatically via the setup command. Config is stored in ~/.devflow/:

# Manual override via environment variable
export DEVFLOW_API_URL=https://api.app.dev-flow.tech
export DEVFLOW_API_TOKEN=your-token

Or create .devflowrc.json in your project or home directory:

{
  "apiUrl": "https://api.app.dev-flow.tech",
  "apiToken": "your-token"
}

You can also start a Runner execution from the DevFlow UI — click the Start Runner button on any flow in the ready or in_progress state.

Key principle: The AI never decides when it’s done. The Runner verifies objectively and only advances when all checks pass.

Both modes use the same pipeline configuration, the same skills, the same gates, and the same DevFlow UI for monitoring.

What You Need

To use the AI agent integration, you need:

1. An AI assistant that supports MCP — such as Claude Code, Cursor, Gemini CLI, or Windsurf
2. The DevFlow MCP server — installed and configured to connect to your DevFlow instance
3. (Optional) The DevFlow Runner — for autonomous flow execution

See MCP Setup for step-by-step installation instructions.