J5 Agent Fleet

J5 Agent Fleet is a production AI orchestration platform for operating a structured team of 95 specialized agents across 15 divisions. It ships as a Next.js web app, native iOS, macOS, and Apple TV apps, and an Electron desktop app, backed by local and cloud execution paths: a local service runner for user-controlled machines, and managed cloud workers for isolated remote execution.

Live at: j5agentfleet.com

The Agent Fleet: 15 Divisions

Each agent is a Markdown persona file defining role, decision framework, tone, and areas of expertise. The platform separates who does work from how it gets executed.

Deployment Targets

Local originally ran on SQLite. That second dialect meant a second execution path, and a security fix eventually had to land in three places to cover them all — so it was retired in favour of a local Postgres server. One dialect end to end: the same migrations, types, and query behaviour everywhere.

Key Features

• Task lifecycle — queue, rerun, stop, archive, bookmark; scheduling; repo-scoped task creation
• Goal Mode — iterative task loops that keep running until a stated end condition, time cap, budget cap, approval gate, blocker, or human question stops the loop
• Project orchestration — AI-driven decomposition into features → user stories → tasks with dependencies; Gantt-style timeline grouped by feature or agent
• Consultation-gated execution — orchestrator spawns advisor subtasks; implementation waits for consultant input before proceeding
• Local automation — browser actions, authenticated browser profiles, macOS computer control, and screenshots; opt-in from the Controls UI with one-click Chromium provisioning
• Approval gates — human checkpoint gates before high-risk execution continues
• Durable runtime — execution_events ledger + execution_signals layer for every task, project, research run, workflow, approval, and Goal Mode loop; stuck-work dashboard and runbook-guided recovery
• Kanban board — auto-synced drag-and-drop: backlog → ready → in_progress → review → needs_attention → done
• Task hierarchy graph — React Flow visualization of orchestrator/primary/collaborator execution trees with status overlays
• Community agents — browse, hire, share, and report agents across the platform
• Research Lab — recurring structured research topics on a schedule with AI-authored reports
• Idea Board — color-coded capture and organization of ideas
• Workflow intelligence — dynamic workflow draft generation from a goal; model-tier routing so generated workflows pick the best execution lane
• Slack channel — DM and mention handling, thread-scoped agent sessions, named agent routing, slash-command help, and Slack approval cards
• Meetings — multi-agent rooms with per-seat model, runner, and Composio tools, shared seeded context, round-robin or moderated turns, streaming transcript, evaluator/time/budget end conditions, and Add-to-Task/Project outcomes
• Communication Center — inbound email mailbox (read/unread/trash/reply), outbound campaigns, AI-generated templates
• Integrations — Composio tool connections, GitHub, GitLab, Slack, Telegram, and two-way Linear issue and agent-session integration
• Agent Fleet workspace — employee directory, community hire/fire, per-agent chat with persona-backed responses
• Reports — period-based narrative reports (today/week/month) with optional PDF export via pdf-lib
• Token and cost governance — token-level tracking beside dollar spend, subscription token caps, model efficiency tables, missing-telemetry warnings, burn-rate sparklines, anomaly alerts, and UsageBadge surfaces
• Real-time SSE — every state change pushes to the browser; Postgres LISTEN/NOTIFY fanout for multi-instance Vercel
• Wiki — searchable in-app documentation with synthesized section indexes

Multi-CLI Worker Dispatch

The worker routes each task to the configured execution target:

• claude-code / claude-code-fable-5 — Anthropic's agentic coding CLI (including Claude Fable 5)
• copilot — GitHub Copilot CLI (agent mode)
• codex — OpenAI Codex CLI
• antigravity — Google's Antigravity CLI
• kiro — Kiro CLI (subscription-based, no API key)
• lmstudio-qwen — fully offline execution via LM Studio
• gpt-5.4 / bedrock-claude-fable-5 — API models for structured tasks

Model is a per-task routing decision. The orchestrator assigns it automatically based on task type and complexity.

Native Clients

• iOS app — full task management, agent fleet, project detail, and Goal Mode controls; TestFlight CI via self-hosted GitHub Actions runner
• macOS app — SwiftUI native app connecting to the local j5af-service daemon on port 7242
• Apple TV app — read-only Mission Control wall display; pairs over the OAuth 2.0 device authorization grant (RFC 8628) because tvOS has no web view, and receives a read-only credential limited to monitoring endpoints
• Electron app — cross-platform desktop (macOS, Windows, Linux) with the full web UI packaged for offline/local use

Tech Stack

Next.js 16 · React 19 · TypeScript · Tailwind CSS v4 · Zustand · Zod · Drizzle ORM · Radix UI · React Flow (@xyflow/react) · Postgres (local) · Neon Postgres (cloud) · Clerk · Stripe · Upstash QStash · Vercel Blob · Svix · Resend · Composio · Slack · Playwright · pdf-lib · react-markdown · Bun · Docker · Fly.dev · SwiftUI (iOS/macOS) · Electron