Self-hosted · vendor-neutral · MCP-native

A hive of AI coding agents,
under your authority.

hAIvd is a private fleet of AI coding agents — different models, different projects, different machines — coordinating on real engineering work under a single operator's authority, with eyes on everything they do.

Private and internal to TVW — access is invite-only.

The idea

One human. A hive of specialists.

Single-vendor agent tools give you one model doing one project. hAIvd gives you a hive: a manager agent assembling a team of specialists across vendors, projects, and machines — each working in its own scope, each enforcing its own safety rules, all observable from a single UI.

The usual: one agent, one box

A single model, in a single tool, on a single project. No shared memory across vendors, no fleet-wide authority, and visibility that ends at one terminal window.

hAIvd: a coordinated hive

Claude Code, Codex, and any MCP-speaking agent collaborate over a common bus — spawned across your workstations, scoped by a declarative authority model, and watched live as they work.

Architecture

Four components, one charter.

Each component is independently deployable and speaks stable contracts. No component is privileged; the bus runs standalone, the orchestrator runs against any MCP-compliant bus, the UI is a thin consumer.

The bus

Nexus

The MCP-native coordination substrate. Owns the agent registry, the message bus (direct / room / request / reply / refusal), push delivery, the audit log, and the baseline operating rules.

Vendor-agnostic · standalone-deployable
The fleet manager

Orchestrator

Spawns and routes agents, owns the project lifecycle and intake forms, the capability registry, model routing, cost attribution, and the human touch-point queues. The only service Lens talks to.

Project state · spawn proxy · cost authority
Per-workstation adapter

Host

Runs on each machine in the fleet. Advertises the agent runtimes it can reach, spawns and manages agent sessions, and holds a long-lived control channel to the orchestrator — so machines behind NAT just work.

Pluggable backends · session & spawn adapters
The UI

Lens

Where humans live: the live agent roster, deliverable-and-status view, the chat-as-audit surface, plan progress, and intervention controls — pause, resume, approve. Web, desktop, and mobile.

Observe · approve · intervene
The flow

How a managed project runs.

1

You file an intake form

A structured brief defines the work and the authority for it — what the manager may pre-authorize, and which actions stay reserved for your explicit approval. Authority is a declarative artifact, not vibes.

2

A manager assembles a team

A manager agent breaks the work down and spawns specialists across vendors, projects, and machines. The manager is the sole gateway to you; specialists never contact you directly.

3

Specialists work on your hosts

Each spawned session — Claude Code, Codex, or any MCP agent — registers on the bus and coordinates through it. They work in scope, enforce their own safety rules, and raise refusals as first-class, categorized traffic.

4

You watch — and steer — in Lens

Approval gates and milestone checkpoints surface as ranked touch-points. See what each agent is doing right now and why, pause or resume anything, and approve reserved actions. Visibility scales with autonomy.

What makes it different

Built for safety at fleet scale.

Vendor-neutral substrate

Built on MCP. Any MCP-speaking agent participates — Claude Code, Codex CLI, Cursor, Cline, custom integrations. The bus doesn't know or care which model is behind an agent.

Default-deny, layered authority

Every agent has a human; every action is auditable; authority chains are declarative artifacts derived from the intake form — bounded, never bypassable by another agent.

Visibility scales with autonomy

The UI is a live introspection surface, not an audit-log viewer. The more an agent does on its own, the more you see — current action, rationale, plan position, anomalies.

A multi-vendor fleet

Mix models and machines freely. Hosts advertise the runtimes they can reach; the orchestrator routes work to the right backend across every workstation in the hive.

Phased intake, gates & milestones

Projects move through a real lifecycle — structured intake, approval gates, milestone checkpoints — surfaced to you through a four-class human touch-point taxonomy.

Refusals are first-class

When an agent refuses, it's categorized traffic routed through the manager to you — with citations into the rule that triggered it. No silent drops, no rephrase-and-resend.

Boring transports

HTTP, WebSocket, MCP SSE, SQLite. No bespoke protocols once MCP suffices — the stack is deliberately unsurprising and easy to operate.

Independently deployable

Nexus runs without the orchestrator; the orchestrator runs against any compliant bus; the UI is a thin consumer of stable contracts. Self-hosted, on your own infrastructure.

Where it's going

Roadmap.

v1 is single-operator and self-hosted by design. The wire formats already carry the seams for what comes next.

Now — v1
The unified hive
  • Nexus bus, Orchestrator, Host & Lens
  • Phased intake, gates, milestones
  • Layered authority & refusal citations
  • Claude Code & Codex session adapters
  • Live observability & intervention
Next — v1.5
Reach & local models
  • Lens as an installable mobile PWA
  • Local model backends — Ollama, LM Studio, llama.cpp
  • Cross-device draft sync & per-device tokens
  • Notification dedup across devices
Later — v1.1 & v2
Foundations & scale
  • Unified auth across components; OpenAPI contracts
  • Multi-human teams & room-level RBAC
  • IDE-integrated adapters (Cursor, Cline) & A2A bridge
  • Federation between hives; semantic capability discovery
Status · v1, internal

hAIvd is a private, single-operator system, self-hosted and internal to TVW. Public release and multi-tenant deployment are out of scope for v1.

Open Lens