AgentWork — Autonomous AI Workbench

Autonomous AI agent orchestration platform for software teams. Hire role-based agents, give them structured workspace memory, let them collaborate in direct chat, mention-based group chat, and autonomous meetings, then turn decisions into project tasks and shipped code.

Website: agentwork.sh

Installation

Requires Node.js 18+.

One-liner install (macOS/Linux)

curl -fsSL https://agentwork.sh/install.sh | bash

The installer bootstraps Node.js if needed, installs the latest agentwork, starts the daemon, and prints the local dashboard URL.

Option 2: npm

npm install -g agentwork
agentwork start

Option 3: From source

git clone https://github.com/agentwork-ai/agentwork.git
cd agentwork
npm install && npm run build && npm link
agentwork start

Open http://localhost:1248 — the onboarding wizard will guide you through setup.

Why AgentWork Is Strong

• Persistent agent workspaces — Every hired agent gets a real workspace with AGENTS.md, ROLE.md, SOUL.md, IDENTITY.md, TOOLS.md, USER.md, HEARTBEAT.md, MEMORY.md, and daily notes under memory/YYYY-MM-DD.md
• Planning-to-delivery loop — Run planning meetings between agents, generate scoped task lists, send them to the Kanban board, and execute them with autonomous agents
• Mixed runtime support — Run agents through API models, Claude CLI, Codex CLI, or provider-auth/OAuth-backed flows from one dashboard
• Built for software organizations — Use built-in leadership and delivery roles, project templates, project manager and main developer defaults, project docs, git automation, and project-aware task execution

Features

Agent System

• Structured Workspace Files — Agents use AGENTS.md, ROLE.md, SOUL.md, IDENTITY.md, TOOLS.md, USER.md, HEARTBEAT.md, MEMORY.md, and daily memory notes
• 22 Built-In Software Roles — Includes Assistant, CEO, CTO, Product Manager, Project Manager, Business Analyst, Engineering Manager, Tech Lead, Frontend/Backend/Full Stack/Mobile Developer, QA, UI/UX, DevOps, Security, Data, ML, Technical Writer, and more
• Role-Specific ROLE.md — Every agent gets a dedicated role contract describing strengths, responsibilities, and expected behavior for that role
• Agent Types — Smart Agents use the full workspace, Worker Agents use only ROLE.md + memory, and CLI Agents skip workspace markdown files
• OAuth / Provider Auth Agents — Hire agents that use stored provider auth from Settings, including Anthropic setup-token, Gemini CLI OAuth, and Codex OAuth, in addition to API key and CLI-auth agents
• Multi-Agent — Hire agents with distinct roles, models, auth methods, and persistent memory
• Agent Cloning — Clone an agent's full configuration and workspace with one click
• Per-Agent Tool Restrictions — Whitelist which tools each agent can use
• Streaming Chat — Token-by-token streaming responses via Socket.io
• Multi-Model Fallback — Automatic fallback to secondary model on API failure
• Context Window Management — Auto-prunes conversation history to prevent truncation
• Vision / Image Support — Agents can analyze images via the read_image tool
• Self-Improving Prompts — Track success/failure rates with prompt improvement suggestions
• Recurring Background Work — HEARTBEAT.md acts as the live checklist for cron-style recurring tasks

Collaboration And Planning

• Direct Agent Chat — Talk to agents one-on-one with streaming responses and memory-aware context
• Group Chat Rooms — Create multi-agent rooms, mention agents with @name, get mention autocomplete, and let only the mentioned agents respond
• Autonomous Meetings — Create topic-driven meetings for a project, invite agents, choose depth (Rapid Alignment, Working Session, Deep Dive), and let agents discuss without user participation
• Task Generation From Meetings — Meetings produce actionable task lists that can be auto-added to the project board or held for user confirmation
• Facilitator-Aware Planning — Leadership roles like CEO, Project Manager, Product Manager, and Business Analyst naturally drive planning and breakdown discussions

Task Management And Delivery

• 5-Column Kanban — Drag tasks to "Doing" and agents execute autonomously
• Flow Tasks — Sequential or parallel multi-step workflows across agents
• Task Dependencies — DAG-based execution — tasks wait for dependencies to complete
• Sub-Tasks — Nested tasks with parent-child relationships
• Task Templates — Save and reuse task blueprints
• Task Labels/Tags — Free-form tagging for filtering and grouping
• Bulk Operations — Multi-select for bulk status change, assign, or delete
• Priority Queue — Auto-start next highest-priority task when one finishes
• Task Scheduling — One-shot timestamps and recurring cron expressions
• Retry Blocked Tasks — Edit description and one-click retry
• Task Comments — Discussion threads separate from execution logs
• Swimlanes — Group Kanban by agent, project, or priority
• Time Estimates & SLA — Track estimated vs actual execution duration

Project Context

• Project Templates — Generate PROJECT.md from built-in templates: Generic, iOS App, Android App, Flutter App, React Native App, Python App, Node API / Service, Next.js Web App, Web Frontend, and Go Service
• Project Manager + Main Developer Fields — Projects can define a planning owner and a default implementation owner
• Smart Task Defaults — New tasks default to the project's Main Developer; meeting-generated tasks fall back to the Project Manager when no owner is suggested
• Project Explorer + Search — Browse files, search code, inspect diffs, and edit inline from the dashboard
• Project-Aware Context — PROJECT.md is used as shared project memory for tasks, chat, worker agents, and CLI agents when relevant

Git Automation (enabled by default)

• Auto Branch — Creates agentwork/<task-slug> branch before each task
• Auto Sync — Pulls latest from main before starting, auto-resolves conflicts
• Auto Commit + PR — Stages, commits, pushes, and opens PR via gh CLI
• Auto Merge — Squash-merges to main (via PR or local merge as fallback)
• Auto Init — Initializes git repo for projects that don't have one

AI Providers

• 9 Providers — Anthropic, OpenAI, OpenRouter, DeepSeek, Mistral, Google, Ollama, LMStudio, Custom
• 80+ Models — Claude 4, GPT-5, Gemini 2.5, DeepSeek V3, Codestral, and more
• Provider Auth In Settings — Store and manage provider auth profiles centrally, then reuse them when hiring agents
• CLI + API + OAuth Auth — Use API keys, local Claude Code / Codex CLI, Anthropic setup-token flows, Gemini CLI OAuth, or Codex OAuth
• Custom Tools — Define your own tools with bash command templates

Security And Ops

• Dashboard Authentication — Optional password protection
• Secrets Encryption — AES-256-GCM encryption for API keys and stored provider auth profiles
• Audit Logging — All CRUD operations logged with timestamps
• Per-Agent Budget Limits — Individual daily spend caps
• Host Filesystem Access — Agents can work outside their own workspace when the runtime and local machine permissions allow it

Integrations

• Telegram Bot — Chat with agents via Telegram
• Slack Bot — DMs and @mentions via Socket Mode
• Discord Bot — Chat with agents via Discord (requires discord.js)
• GitHub — Auto-create issues from blocked tasks, PRs from completed tasks
• Linear / Jira — Bidirectional task sync
• Email Notifications — SMTP-based alerts for task completion/blocked
• Webhooks — Inbound POST /api/webhooks/trigger for CI/CD
• MCP Server — Expose tasks/agents/projects to Claude Desktop
• VS Code Extension — Status, task creation, agent listing from editor
• Plugin System — Third-party tools, hooks, and integrations

UI/UX

• Dark/Light Mode — System-aware with 6 accent color presets
• Markdown Rendering — Rich formatting in chat and task details
• Syntax-Highlighted Code Viewer — JS/TS/Python/JSON/CSS/Markdown
• In-Line File Editor — Edit and save files from the dashboard
• Diff Viewer — Unified diff with color-coded added/removed lines
• Keyboard Shortcuts — Cmd+K, Cmd+1-7, ? for help overlay
• Execution Log Filtering — Type filters, text search, color-coding
• Live Activity Feed — Real-time event stream on dashboard
• Onboarding Wizard — Guided 4-step setup for first-time users
• Responsive Mobile Layout — Hamburger menu, compact bottom bar
• PWA Support — Installable as mobile app with service worker
• Analytics Dashboard — Spend charts, agent utilization, model comparison

Architecture

Node.js daemon (port 1248)
├── Express REST API         Tasks, agents, projects, rooms, meetings, settings, tools
├── Socket.io                Real-time task updates, execution logs, chat, meetings, streaming
├── Next.js 14 SSR           Dashboard UI (React 18, Tailwind CSS)
├── SQLite (better-sqlite3)  Local database with versioned migrations
└── Services
    ├── executor.js          Task execution with concurrency queue, git automation
    ├── scheduler.js         Cron and one-shot task triggers
    ├── ai.js                Multi-provider completion + streaming engine
    ├── agent-context.js     Agent workspace loading, role logic, prompt assembly
    ├── meetings.js          Autonomous meeting orchestration and task generation
    ├── provider-auth.js     Provider auth storage, OAuth handoff, CLI auth sync
    ├── platforms.js         Telegram, Slack, Discord bots
    ├── rag.js               Keyword-based file retrieval for project context
    ├── github.js            GitHub issue/PR integration
    ├── email.js             SMTP notification service
    ├── plugins.js           Plugin system loader
    └── crypto.js            AES-256-GCM encryption for API keys

CLI

agentwork start [-p PORT] [-f]    # Start daemon (default port 1248, -f for foreground)
agentwork restart [-p PORT] [-f]  # Restart daemon and preserve the previous port by default
agentwork stop                    # Graceful shutdown
agentwork update [--tag TAG]      # Install the latest npm release and restart if currently running
agentwork status                  # PID, URL, active agents
agentwork logs [-n N] [-f]        # Tail server logs
agentwork clean                   # Clear temp files and logs
agentwork task list [--status S]  # List tasks with optional filter
agentwork task create <title>     # Create task (--description, --priority, --agent)
agentwork agent list              # List all agents with status

To install the CLI globally: npm link

Dashboard Pages

| Page | Path | Description | |------|------|-------------| | Dashboard | / | Stats, live activity feed, task pipeline, quick actions | | Projects | /projects | Project CRUD, templates, project manager/main developer defaults, file explorer, search, inline editor | | Tasks | /kanban | 5-column Kanban with swimlanes, bulk ops, dependencies, templates | | Chat | /chat | Direct chat plus mention-based group chat rooms with autocomplete and room editing | | Meetings | /meetings | Autonomous planning meetings that turn discussions into project task lists | | Office | /office | Live agent telemetry + execution timeline / Gantt chart | | Analytics | /analytics | Spend charts, agent performance, model comparison | | Agents | /agents | Hire/fire/clone agents, choose role and agent type, edit workspace files, configure auth mode | | Settings | /settings | API keys, provider auth/OAuth, budget, git behavior, security, templates, preferences |

AI Providers

| Provider | Auth | Example Models / Runtime | |----------|------|--------------------------| | Anthropic | API key or setup-token | claude-opus-4-1, claude-sonnet, claude-haiku | | OpenAI | API key | gpt-5, gpt-4o, o3, o4-mini | | OpenAI Codex | OAuth / provider auth | Codex CLI runtime and ChatGPT Codex backend | | Google | API key or Gemini CLI OAuth | gemini-2.5-pro, gemini-2.5-flash | | OpenRouter | API key | 200+ models | | DeepSeek | API key | deepseek-chat (V3), deepseek-reasoner (R1) | | Mistral | API key | mistral-large, codestral | | Ollama / LMStudio | Custom URL | Any local model | | Claude Code | CLI auth | Local Claude CLI runtime |

Task Execution

1. Create a task on the Kanban board and assign an agent + project
2. Drag the task to Doing (or set a schedule/cron trigger)
3. Git automation kicks in: sync from main → create feature branch
4. The agent loads context based on agent type: Smart Agent = full workspace, Worker Agent = ROLE.md + memory, CLI Agent = no workspace markdown; PROJECT.md is added when project context is attached
5. Execution logs stream to the UI in real time (thoughts, commands, file changes)
6. On success: auto-commit → push → PR → merge to main → task moves to Done
7. On failure: task moves to Blocked → retry with modified description
8. Next queued task for the same agent auto-starts

Flow tasks chain multiple steps across different agents, sequentially or in parallel.

Data Directory

~/.agentwork/
├── db/agentwork.db           # SQLite database (encrypted API keys)
├── agents/<id>/              # Per-agent memory
│   ├── AGENTS.md             # Operating rules and learned conventions
│   ├── ROLE.md               # Role-specific skills, strengths, and default responsibilities
│   ├── SOUL.md               # Persona, tone, and boundaries
│   ├── IDENTITY.md           # Name, vibe, emoji, avatar, identity hints
│   ├── TOOLS.md              # Local environment notes and setup specifics
│   ├── USER.md               # User preferences and profile
│   ├── HEARTBEAT.md          # Checklist for recurring scheduled work
│   ├── MEMORY.md             # Curated long-term memory
│   ├── memory/YYYY-MM-DD.md  # Daily raw notes and recent context
│   └── *.md                  # Custom memory files
├── TEAM.md                   # Shared memory across all agents
├── plugins/                  # Third-party plugins
├── logs/agentwork.log        # Server logs
└── agentwork.pid             # Daemon PID

Agent type behavior:

• Smart Agent: loads all workspace markdown files
• Worker Agent: loads ROLE.md, MEMORY.md, and recent daily memory notes
• CLI Agent: loads no workspace markdown files; only PROJECT.md is used when project context is attached

Environment Variables

| Variable | Default | Description | |----------|---------|-------------| | PORT | 1248 | Server port | | NODE_ENV | development | production for optimized builds | | AGENTWORK_DATA | ~/.agentwork | Data directory path | | AGENTWORK_ROOT | Auto-detected | Project source root | | AGENTWORK_SETTING_* | — | Override any setting (e.g., AGENTWORK_SETTING_ANTHROPIC_API_KEY) |

Development

npm run dev           # Dev server with hot reload
npm run build         # Production build
npm start             # Production server

API Documentation

GET /api/docs returns a JSON listing of all 50+ API endpoints.

Key endpoints:

• /api/tasks — Task CRUD, bulk operations, subtasks, comments, replay
• /api/agents — Agent CRUD, clone, metrics, prompt analysis, inbox
• /api/projects — Project CRUD, templates, file search, git status, diff, health score
• /api/meetings — Meeting drafts, autonomous runs, transcripts, and task application
• /api/settings — Settings, provider auth, budget, cost breakdown, reports, export, audit logs
• /api/templates — Task template CRUD
• /api/tools — Custom tool CRUD
• /api/rooms — Group chat rooms
• /api/webhooks/trigger — External task trigger
• /api/health — System health check

Tech Stack

Next.js 14 · React 18 · Tailwind CSS · Express · Socket.io · SQLite · Anthropic SDK · Claude Agent SDK · OpenAI SDK · Codex SDK · Telegraf · Slack Bolt · node-cron · Framer Motion · dnd-kit

License

MIT