NOTE: This is still in the prototype phase and not ready for use

Fort

Fort is a task-centric personal AI agent platform. It is not a chat wrapper. Every interaction creates a task with clear ownership and status. Deterministic services handle orchestration, memory, scheduling, and reflection. Users create long-lived specialist agents through a web portal, each defined by a SOUL.md that shapes its personality and goals.

Quick Start

# Install globally from npm
npm install -g @fort-ai/cli

# Initialize Fort and open the portal wizard
fort init

# The portal opens at http://localhost:4077
# Walk through the setup wizard to create your first agent

Use as a library

npm install @fort-ai/core

import { Fort } from '@fort-ai/core';

Development

# Clone the repo
git clone https://github.com/fort-ai/fort.git
cd fort

# Install dependencies
npm install

# Build
npm run build

# Link the CLI globally
npm link --workspace=packages/cli

Architecture

Fort separates services from agents.

Services are deterministic infrastructure. They do not use LLMs unless they hit ambiguity. They live in packages/core/src/services/.

Service	Role
Orchestrator	Creates a task for every chat message and routes it to the correct agent
Memory	Manages the graph-based knowledge store
Scheduler	Executes cron-scheduled routines
Reflection	Periodically scans completed tasks for missed action items, creates follow-up tasks

Agents are user-created specialists. Each agent has a SOUL.md that defines its personality, goals, rules, voice, and boundaries. The SOUL.md is injected into the LLM system prompt at runtime.

Chat messages go to the user's default agent (or a selected long-lived agent). The orchestrator service creates the task and routes it -- it does not answer questions itself.

Task-Centric Design

Every chat message creates a task. Tasks have:

• result -- the outcome
• assignedTo -- 'agent' or 'user'
• Status -- To Do, In Progress, Done

The portal at http://localhost:4077 shows a kanban board with clear ownership across these columns.

Agent Lifecycle

Users create agents through the portal setup wizard or the CLI. The wizard collects name, goals, and emoji, then generates a SOUL.md.

Each agent gets a directory:

.fort/data/agents/<agent-slug>/
  identity.yaml   # name, status, metadata
  SOUL.md          # personality, goals, rules, voice, boundaries
  tools/           # agent-specific tool configs

Lifecycle commands: fort agents create, fork, retire, revive.

CLI Reference

Command	Description
fort init	Initialize Fort, open portal wizard
fort doctor	Health checks across all modules
fort status	System status overview
fort agents list	List specialist agents
fort agents create	Create a specialist agent
fort agents fork <id>	Clone an agent with a new name
fort agents retire <id>	Deactivate an agent
fort agents revive <id>	Reactivate a retired agent
fort agents inspect <id>	Deep-inspect an agent
fort tasks	Show task graph (kanban view)
fort memory stats	Memory graph statistics
fort tools list	Tool registry
fort tokens	Token usage and cost
fort tokens budget-set	Set spending limits
fort schedule list	Cron jobs
fort routines list	Scheduled routines
fort flags list	Feature flags
fort plugins list	Loaded plugins
fort harness start	Start a self-coding cycle
fort rewind	Snapshot management
fort introspect	System profile and capability map
fort llm setup	Authenticate with Claude
fort llm status	LLM config and usage stats
fort llm ask <prompt>	Send a prompt (--model fast|standard|powerful)

Project Structure

packages/
  core/             TypeScript core
    src/
      services/       Orchestrator, Memory, Scheduler, Reflection
      agents/         Agent framework, SOUL.md handling
      task-graph/     Task tracking and kanban state
      memory/         Graph-based memory store
      module-bus/     Event-driven message bus
      llm/            Anthropic Claude API client
      threads/        Conversation threading
      tools/          Tool registry
      permissions/    Tiered action model
      flows/          Deterministic workflow engine
      tokens/         Usage tracking and budgets
      scheduler/      Cron scheduling
      routines/       Cron-scheduled routines
      diagnostics/    FortDoctor health checks
      harness/        Self-coding cycle
      integrations/   Gmail, Calendar, iMessage, Brave Search
      ipc/            WebSocket server
      plugins/        Plugin system
      rewind/         Snapshot backup and restore
      specs/          Spec-driven development
  cli/              Commander.js CLI
  swift-shell/      macOS menu bar app (AppKit + WebSocket)
  dashboard/        Tauri + React dashboard and portal (http://localhost:4077)
  docs/             Docusaurus documentation

Tech Stack

• Runtime: Node.js / TypeScript
• Tests: Vitest
• CLI: Commander.js
• DB: better-sqlite3 (SQLite per module)
• Config: YAML + JSON Schema (ajv)
• LLM: Anthropic Claude (Haiku/Sonnet/Opus tiered routing)
• macOS: Swift/AppKit menu bar + Tauri dashboard
• Docs: Docusaurus

License

MIT