Agently.ai — Agent Management Platform

A local-first web platform for creating, configuring, and orchestrating AI agents powered by Amazon Bedrock (Claude). Work is managed through a Kanban board: drag a task to In Progress and the assigned agent runs with streaming output, conversation history, and human-in-the-loop review.

Built as a single-user developer tool — no cloud deployment required beyond AWS Bedrock access.

---

Problem

Running multiple specialized AI agents (code review, bugfix, research, etc.) from the terminal does not scale: there is no shared task queue, no visibility into runs, and no structured review loop. Agently.ai wraps agent execution in a Kanban workflow so you can assign agents to tasks, watch output stream in real time, leave feedback, and re-run with or without prior context.

---

Features

Implemented

• Kanban board — fixed columns: Backlog → To Do → In Progress → Review → Done
• Drag-and-drop tasks — moving a task to In Progress triggers agent execution
• Agent CRUD — create agents with name, description, Bedrock model, and CLAUDE.md system prompt
• Streaming execution — Server-Sent Events (SSE) from /api/execute with live output in a dialog
• Conversation modes — continue (full history replay) or fresh (start over)
• Human-in-the-loop — review output, add feedback, re-run from Review
• Execution history — per-task run log with status, timestamps, and output
• File-based persistence — agents and tasks stored as JSON under data/

Planned

• Teams — sequential, parallel, and orchestrated multi-agent workflows
• Skills marketplace — install skills from anthropics/claude-plugins-official or create custom ones
• Docker sandbox — isolated agent execution via containerized Claude Code CLI
• AI-assisted setup — generate agent prompts and team structures via Bedrock

---

Architecture

┌─────────────────────────────────────────────────────────────┐
│                     Next.js App (React 19)                  │
│  ┌──────────┐  ┌─────────────────────────────────────────┐  │
│  │ Sidebar  │  │              Kanban Board               │  │
│  │ Agents   │  │  Backlog │ To Do │ In Progress │ ...   │  │
│  └──────────┘  └─────────────────────────────────────────┘  │
└──────────────────────────┬──────────────────────────────────┘
                           │ REST + SSE
┌──────────────────────────▼──────────────────────────────────┐
│                    API Routes (App Router)                  │
│  /api/tasks   /api/agents   /api/execute (SSE stream)       │
└──────────────────────────┬──────────────────────────────────┘
                           │
         ┌─────────────────┼─────────────────┐
         ▼                 ▼                 ▼
   data/tasks/      data/agents/     @anthropic-ai/bedrock-sdk
   (JSON files)     (config +        (Claude via AWS Bedrock)
                     CLAUDE.md)

Execution flow

1. User assigns an agent to a task and drags it to In Progress.
2. Frontend opens an output dialog and connects to GET /api/execute?task_id=….
3. Backend loads task + agent config, builds the message array (with optional history), and streams from Bedrock.
4. Tokens are pushed over SSE; the result is persisted to data/tasks/<id>.json.
5. User reviews output in Review, approves to Done, or re-runs with feedback.

---

Tech Stack

Layer	Technology
Framework	Next.js 16 (App Router)
UI	React 19, Tailwind CSS 4, shadcn/ui, Base UI
Drag & drop	@dnd-kit
Data fetching	SWR
AI	@anthropic-ai/bedrock-sdk → Amazon Bedrock (Claude Opus / Sonnet / Haiku)
Persistence	Local JSON files (data/)
Container (WIP)	Docker + dockerode + xterm.js

---

Prerequisites

• Node.js 20+
• AWS account with Bedrock access and Claude models enabled in your region
• AWS CLI profile configured (see Environment variables)

---

Setup

# Clone
git clone https://github.com/EdoardoMorucci/agent-platform.git
cd agent-platform

# Install dependencies
npm install

# Configure environment
cp .env.example .env.local
# Edit .env.local with your AWS profile and region

# Start dev server
npm run dev

Open http://localhost:3000.

AWS Bedrock

1. Enable Claude models in the Amazon Bedrock console for your region (default: eu-west-1).
2. Create a named AWS profile (e.g. bedrock) in ~/.aws/credentials:

[bedrock]
aws_access_key_id = YOUR_ACCESS_KEY
aws_secret_access_key = YOUR_SECRET_KEY

Alternatively, use AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY environment variables or an IAM role.

Application defaults live in config.json:

{
  "aws_profile": "bedrock",
  "aws_region": "eu-west-1",
  "default_model": "eu.anthropic.claude-sonnet-4-6",
  "claude_cli_path": "claude"
}

---

Environment Variables

Variable	Required	Default	Description
AWS_PROFILE	Yes*	bedrock	Named AWS profile for Bedrock credentials
AWS_REGION	Yes	eu-west-1	AWS region where Bedrock models are enabled
PORT	No	3000	Dev server port

* Or use standard AWS credential env vars / IAM role instead of a profile.

Copy .env.example to .env.local and adjust values for your setup.

---

Docker (Work in Progress)

A base image for running Claude Code CLI inside a container is included for upcoming sandboxed execution:

# Build the agent image
docker build -t agently-claude -f docker/Dockerfile docker/

# Run interactively (mount your workspace)
docker run -it --rm \
  -v "$(pwd)/workspace:/workspace" \
  -v "$HOME/.aws:/root/.aws:ro" \
  -e AWS_PROFILE=bedrock \
  -e AWS_REGION=eu-west-1 \
  -e CLAUDE_CODE_USE_BEDROCK=1 \
  agently-claude

Full Docker-based execution with PTY terminal streaming is planned on the docker-implementation branch.

---

Project Structure

agent-platform/
├── src/
│   ├── app/
│   │   ├── api/
│   │   │   ├── agents/          # Agent CRUD + CLAUDE.md editor
│   │   │   ├── tasks/           # Task CRUD
│   │   │   └── execute/         # SSE streaming execution endpoint
│   │   ├── layout.tsx
│   │   └── page.tsx             # Kanban view
│   ├── components/
│   │   ├── agents/              # Agent panel, form, output dialog
│   │   ├── kanban/              # Board, columns, task cards
│   │   ├── layout/              # Shell, sidebar, top bar
│   │   └── ui/                  # shadcn primitives
│   ├── hooks/                   # SWR hooks for agents & tasks
│   └── lib/                     # Types, file I/O helpers
├── data/
│   ├── agents/<id>/             # config.json + CLAUDE.md per agent
│   ├── tasks/                   # One JSON file per task
│   ├── teams/                   # (planned)
│   └── skills/                  # (planned)
├── docker/
│   └── Dockerfile               # Claude Code CLI container image
├── docs/superpowers/            # Design specs and implementation plans
├── config.json                  # App defaults (AWS profile, region, model)
├── .env.example
└── package.json

---

Available Bedrock Models

UI Label	Model ID
Claude Opus 4.6	eu.anthropic.claude-opus-4-6-v1
Claude Sonnet 4.6	eu.anthropic.claude-sonnet-4-6
Claude Haiku 4.5	eu.anthropic.claude-haiku-4-5-20251001

Model IDs use EU cross-region inference profiles. Adjust in src/lib/types.ts if your region differs.

---

Scripts

Command	Description
npm run dev	Start development server
npm run build	Production build
npm run start	Start production server
npm run lint	Run ESLint

---

License

No license file is included yet. All rights reserved unless otherwise specified.