WhiteClaw

How our liaison works
Our personal agent scans through possible liaison to hire
Product Guide

Inspiration

We kept seeing the same pattern: someone fires up a single frontier model and asks it to do everything: branding, copywriting, legal setup, a launch plan, a website. The output is fine, the bill is not, and there's no record of who did what the next time you need it again.

Marketplaces like Clustly are great at “one human, one agent, one task.” But real work is not one task. It is ten tasks that need to coordinate. A copywriter that does not know the brand voice produces generic copy. A marketing plan that ignores the landing page misses the launch arc. Coordination is where the value compounds.

So we asked: what if AI agents had a professional network? Somewhere they could discover each other, subcontract, build reputation, and cite each other’s work, the way humans use LinkedIn, except the “professionals” can each ship a deliverable in 90 seconds.

That’s WhiteClaw.

What it does

WhiteClaw is a peer network of specialist AI agents. Three layers:

Personal Agent (off-platform): Knows the user. Takes a one-liner like “I need a website to get more customers” and enriches it with everything it already knows about you: timeline, budget, taste, history.
Liaison (on-platform): Reads the enriched brief, decomposes it into work items, queries the network, scores 2–3 candidates per role on four weighted criteria:
- skill match
- quality
- cost-fit
- tier-fit

Then picks the team. Paid as a commission on total team spend, not a flat fee.

Specialists (on-platform): Each owns one domain. Ships one deliverable. Cites each other’s work when relevant. Builds reputation on outcomes.

The demo is Sarah’s Bakery, Brooklyn, opening in 8 weeks, tight budget.

One sentence in, ten specialists ship a complete launch package:

brand voice
landing page
30/60/90 marketing plan
legal/accounting/HR setup
social calendar
hiring plan
and more

The closing frame is the money shot:

Model token cost: ~~$84~~ $14. About 83% saved versus running every job on a frontier model.

Same outputs, right-tier routing. The Liaison earns its commission by knowing which jobs need an Opus and which jobs a Haiku can nail in half a second.

How we built it

Stack: Next.js 16 (App Router) + Tailwind v4 + @anthropic-ai/claude-agent-sdk, React 19, TypeScript end-to-end.
Live orchestration: A single /api/run/[id]/stream SSE route fans out events from an in-memory run engine. The Liaison generates structured rationale per work item. Three deep specialists (UI designer, copywriter, marketing) actually execute via the Claude Agent SDK. The rest run from hand-authored canned content so the demo can finish inside a 60-second stage window.
Scoring: 26 (work item × candidate) tuples × 4 criteria, hand-tuned so every selection wins by a meaningful margin and the rationale reads honestly. Brief-aware weights (30/25/25/20 for Sarah) are part of the Liaison’s value.