Open Source | MIT Licensed

Open-source orchestration for edge-native AI fleets

Deploy AI agents from cloud to edge. Hub coordinates. Spokes execute. Your Tailnet is the trust boundary.

terminal
$ pnpm seaclip onboard

Up and running in under 2 minutes

Three steps to your fleet

01

Define the mission

Set company goals cascading to projects and tasks. Every agent knows the objective.

02

Deploy the fleet

Register edge devices, assign agents — cloud and local. Your Tailnet defines the boundary.

03

Monitor and govern

Dashboard, approvals, cost control, live telemetry. You're the board.

Everything you need to run AI at the edge

BYOA (7 adapters)

Claude, Ollama, SeaClaw, Codex, Process, HTTP, Telegram

Goal Alignment

Hierarchical goals cascade from company to agent

Heartbeats

5-field cron scheduling per agent

Cost Control

Per-run token and cost attribution

Multi-Company

Isolated tenants sharing one instance

Ticket System

Structured issues with atomic checkout

NEW

Edge Mesh

Device fleet with health scoring

NEW

Hub Federation

Multi-hub sync and task routing

NEW

Spoke Thin-Client

Mobile-friendly device UI

Secrets Management

AES-256-GCM encrypted store

Claude Code Adapter

Native git/branch/PR integration

WebSocket Events

Real-time dashboard updates

Seven adapters, one protocol

Connect any AI backend. Cloud or edge, every adapter speaks the same language.

Claude
Ollama
SeaClaw
Codex
Process
HTTP
Telegram
Hub CEO
Cloud Engineers Claude, Codex, HTTP
Spoke Engineers Ollama, SeaClaw, Process

Goals flow from hub to spoke

Every agent knows its mission. Hierarchical goals cascade from company objectives down to individual tasks.

Company Goal Increase uptime to 99.9%
Project Deploy monitoring agents
Task Configure edge heartbeats
Agent Assignment Local-Analyst on Pi-cluster-01

Every pulse, accounted for

5-field cron scheduling per agent. Edge device telemetry side by side.

Agent Heartbeats

Local-Analyst */5 * * * *
2s ago
Cloud-Writer 0 */2 * * *
45s ago
Edge-Monitor */10 * * * *
8s ago
Backup-Agent 0 0 * * *
3h ago

Device Telemetry

pi-cluster-01
CPU 45%
Memory 62%
GPU 38%
Temp 52°C
Disk 71%

Every token has a price tag

Per-run attribution. Know exactly what each agent costs, split by edge vs cloud.

Total Monthly Cost $127.43
Cloud 70%
$89.21
Edge 30%
$38.22

Top Agents by Cost

Cloud-Writer $45.30
Cloud-Analyst $43.91
Local-Analyst $12.10

One instance, many worlds

Isolated tenants sharing infrastructure. Each company sees only its own agents, devices, and data.

EdgeCorp IoT
8 Agents
12 Devices
Research Lab Alpha
4 Agents
3 Devices
Home Automation
6 Agents
24 Devices

Complete data isolation. Shared compute.

Structured work, from edge to merge

Every task flows through a spoke workflow with full traceability.

pending
assigned
in_progress
pr_raised
merged
done
#42 in_progress

Deploy monitoring agent to pi-cluster-02

Assignee Local-Analyst
Branch feat/monitor-42
Created 2h ago
Assigned to Local-Analyst 1h ago
Branch created 45m ago
PR raised 10m ago
SEACLIP EXCLUSIVE

Your devices, orchestrated

Register any device. Monitor everything. Auto-reassign on failure.

Hub
Pi-01
Jetson-02
Mac-Studio
Pi-03
Phone-01

Device Registration

POST /api/devices — token-based onboarding

Live Telemetry

CPU, RAM, GPU, temp, disk at 10-30s intervals

Health Scoring

0-100 scale, automatic threshold alerts

Auto-Reassignment

Unhealthy device? Tasks move to healthy agents

Supported devices

Raspberry Pi  |  Jetson  |  Mac  |  Linux  |  Android  |  iPhone

SEACLIP EXCLUSIVE

Hubs that talk to hubs

Multi-hub sync protocol. Cross-hub task routing. Geographic distribution.

Hub US-West 12 agents
Hub EU-Central 8 agents
Hub AP-Tokyo 6 agents

Mutual Handshake

Public key exchange, verified identity

Event-Driven Sync

Real-time updates + 5-min reconciliation

Cross-Hub Routing

Route tasks to the best hub for the job

Conflict Resolution

Last-writer-wins with 30-day tombstones

Built in the open

Extensible

Seven adapter types. Add your own in minutes. Every integration follows the same protocol.

Adaptable

PGlite for zero-config, PostgreSQL for scale. Local or cloud. Your infrastructure, your rules.

Open Source

MIT licensed. Fork it, extend it, self-host it. No vendor lock-in, no phone-home.

Pre-built fleet configurations

Coming soon — one-click deployment templates for common edge scenarios.

Coming Soon

Smart Home Fleet

Manage home automation agents across Pis and phones

Coming Soon

Industrial IoT Monitor

Factory floor sensor agents with health scoring

Coming Soon

Dev Agency

AI coding agents with PR workflow and cost tracking

Coming Soon

Research Lab

GPU cluster agents for ML training and inference

You're the board

Every agent action can require approval. Set policies, review decisions, maintain control.

Approval Request

Agent Cloud-Writer requests approval to:

Merge PR #38 into main

Active Policies

  • PRs require human approval
  • Cost > $10 requires review
  • Edge deployments need device-owner sign-off

More than an orchestrator

An AI fleet commander

Coordinate dozens of agents across cloud and edge from a single dashboard.

A device fleet manager

Register, monitor, and health-score every device in your mesh.

A privacy-first platform

Your data stays on your network. Local agents never leave your Tailnet.

A goal-driven system

Every agent action traces back to a company objective.

A governance layer

Approvals, cost limits, and audit trails for every decision.

A developer tool

Git-native PR workflows, Claude Code integration, and a CLI that feels like home.

Sound familiar?

  1. 01

    “My agents have no structure”

    Hierarchical goals, projects, and tasks give every agent a mission.

  2. 02

    “I can't track what agents are doing”

    Real-time heartbeats, audit logs, and WebSocket events.

  3. 03

    “Agent costs are a black box”

    Per-run token attribution with budget enforcement.

  4. 04

    “I need human oversight”

    Typed approval workflows for any agent action.

  5. 05

    “I have multiple teams”

    Multi-company isolation on a single instance.

  6. 06

    “My agents need different backends”

    Seven adapters: Claude, Ollama, SeaClaw, Codex, Process, HTTP, Telegram.

  7. 07

    “My edge devices run agents but I have no visibility”

    Edge mesh gives you real-time telemetry, health scores, and auto-reassignment. SeaClip Exclusive

  8. 08

    “I can't trust cloud APIs with sensitive data”

    Environment isolation keeps local agents on your Tailnet. Cloud agents stay in the cloud. SeaClip Exclusive

Built for builders

Express 5

Async-first API with typed error handling

Drizzle ORM

Type-safe schema, zero runtime overhead

PGlite / PostgreSQL

Embedded for dev, Postgres for prod

React 19 + Vite 6

Dashboard with TanStack Query and Radix UI

WebSocket

Live events pushed to every connected client

Commander.js CLI

Full CLI with onboard wizard and health checks

AES-256-GCM

Secrets encrypted at rest with scrypt key derivation

Edge-native telemetry

10-30s device reporting with health scoring

Hub federation protocol

Event-driven sync with conflict resolution

Environment routing guards

Cloud and local agents never cross boundaries

Clear boundaries

Not a chatbot

A fleet orchestrator that manages agents, not conversations.

Not a model provider

Bring your own models. We connect them.

Not a cloud-only platform

SeaClip runs where your data lives. Edge devices, air-gapped networks, home labs.

Not a no-code tool

Built for developers. CLI-first, Git-native, fully extensible.

Not a monitoring dashboard

Monitoring is one feature. Orchestration is the mission.

Not Paperclip

Forked with love, re-architected for the edge. Hub-spoke, device mesh, environment isolation.

Frequently asked questions

How is SeaClip different from Paperclip?

SeaClip is a fork of Paperclip, re-architected for edge-native deployment. We added hub-spoke architecture, edge device mesh, environment isolation, device telemetry, health scoring, hub federation, and a spoke thin-client. The core orchestration DNA is shared — everything else is built for the edge.

What devices can run as spokes?

Anything that runs Linux or macOS: Raspberry Pi, Jetson boards, Mac, Linux servers, and even phones. The SeaClaw C binary is just 709KB and runs on minimal hardware.

Do I need Tailscale?

Recommended but not required. Tailscale provides zero-config mesh networking and is the default trust boundary for environment isolation. Without it, you can still use SeaClip with any VPN or direct network access.

Can spokes run without internet?

Yes. Local agents (Ollama, SeaClaw, Process) work fully offline. Cloud agents (Claude, HTTP) need connectivity. Environment isolation ensures local agents never attempt external calls.

How does environment isolation work?

Agents are tagged as 'cloud' or 'local'. Routing guards enforce that local agents only communicate within your network, and cloud agents only reach external APIs. The boundary is your Tailnet.

What's the SeaClaw binary?

A 709KB C11 binary that runs on edge devices as a lightweight agent runtime. It reports telemetry, executes tasks, and communicates with the hub over HTTP.

Is SeaClip production-ready?

SeaClip is in active development. The core orchestration, edge mesh, and hub federation are functional. We recommend it for development, testing, and home lab use. Production hardening is ongoing.

How do I contribute?

Fork the repo, pick an issue, and open a PR. We use conventional commits and require passing typecheck and tests. See CLAUDE.md in the repo for full coding conventions.

Ready to deploy your fleet?

Open source. MIT licensed. Run it anywhere.