CampaignLoop live ↗

Idea

CampaignLoop is a single source of truth for a marketing team that works with AI agents. Each team member runs their own agent on their own desktop, for example Claude Code, and the agents connect over MCP to one central system. The central system holds the tasks, the permissions and the approvals. It is a hybrid: an AI system that runs itself in its own cloud, and a set of local, distributed agents that cooperate with it.

Problem

When every person on a team has a personal agent, the agents drift apart. Two of them start the same work, one overwrites the other, and nobody knows who approved a change that spends money. The team needs one place that decides who works on what, and a record of every decision.

Approach

• Control plane. One workspace in the cloud owns the state: campaigns, tasks, permissions, approvals and an audit log. Postgres is the real synchronisation layer; MCP is only the interface to it.
• Local agents on the edge. Each member’s own agent asks the MCP gateway for the next task, does the creative or browser-based work locally, and submits the result back.
• Task claiming. An agent claims a task, and the system locks it with a lease, so nobody else starts the same work. Versioned entities and optimistic locking stop silent overwrites.
• Human gates. Every spend, launch and targeting change needs human approval. The optimizer may only pause weak ads and file budget proposals.

flowchart TD
  A1[Member agent, local desktop] --> G[MCP gateway]
  A2[Member agent, local desktop] --> G
  G --> C[Control plane: tasks, leases, permissions, approvals, audit]
  C --> D[(Postgres)]
  C --> M[Channel adapter: Meta]
  C --> H{{Human approval}}

The first workflow on top of it is marketing: brief, strategy, plan, creatives, a campaign created paused, launch, monitoring and optimization. Meta is the first channel adapter, and an offline demo mode exists.

Stack

Next.js 16, React 19, TypeScript, Postgres with Drizzle, Tailwind 4, NextAuth v5, Vitest, Docker, GitHub Actions, and the Vercel AI SDK through OpenRouter.

Status

The task queue (list, claim, submit) and the approval gates are implemented. Concurrent multi-user access through the gateway is the next step, and the product is unfinished. The repository is private for now. I am happy to walk through it on request.