subagent*.com · workers launchpad application · 2026-07-02

The content network built for the web's next reader: AI agents.

94 production domains, operated end-to-end by 20 Claude coworkers under human approval gates. One founder. Two line items of cost. All of it on Cloudflare Workers.

$ curl -s https://subagentdomains.com/llms.txt | head -2
# subagentdomains.com — the live portfolio index
> 50 Cloudflare Workers on one account…

Every number in this deck is live and curl-verifiable · ↓/↑ to navigate

why now · three dated facts

Agent traffic became meterable this month.

Almost nothing on the web is built to be found, consumed, cited, and paid for by agents. Our whole network is. Whoever holds agent-consumable inventory when per-crawl payment goes GA owns a new revenue class on day one.

product

An agent can find us, use us, and cite us — in four requests.

Positioning, kept honest: the user is an agent — we optimize its cost-to-value. The buyer is the agent's operator, and we meet them at Cloudflare's meter. Two different actors; the deck never blurs them.

GET  subagentroles.com/llms.txt            → family index, API routes
GET  subagentapi.com/.well-known/api-catalog → 49 sites, machine-readable
POST subagentroles.com/mcp {initialize}      → protocolVersion 2025-06-18
POST /mcp {tools/call get_roles}             → 18 rows, each with cite_as:
  "https://subagentroles.com/api/roles/role_agent_a2a-bridge"
# that cite_as now lives in the agent's memory → next agent follows it in

Run it yourself: subagentapi.com/api/catalog · subagentcontracts.com/api/changes

Support is a product property: agents self-serve — llms.txt + MCP docs on every zone, errors as typed JSON. Our first ticket was filed for an agent: our own Python script, 403'd by our own bot-protection while llms.txt invited it in (fleet scripts now send identified UAs; the zone policy is the operator's call — tsk_0536fa28-a30).

live today · measured 2026-07-02 · scripts/flywheel-metrics.sh

Not a prototype — a fleet, with momentum you can diff.

Traction is real but no longer single-day velocity — the fleet has grown from 50 to 94 domains and 20 coworkers since this deck's 2026-07-02 origin, verified live 2026-07-14, not re-measured same-day. Every metric here carries provenance the same way: subagentcache.com's source column never lets a hand-pasted number pass as a live API feed — investors can diff the claim, not just read it. Directory: subagentdomains.com

growth model

A flywheel that's shipped, not sketched.

North star: returning agent visits — change-feed re-polls and cite_as re-entries — because a returning agent is proof of extracted value, not curiosity. Endpoint and catalog counts are input levers; the same script that produced slide 04 measures them before and after every lever. A scoreboard, not a narrative.

Grounded in Cloudflare's docs-for-agents pattern and Anthropic's agent memory + webhook models.

monetization · cloudflare-native

Traffic first. The Gateway meters it — no payment code on our side.

• Monetization Gateway prices any page, API route, or MCP tool at the zone edge; our Workers stay payment-free. Settlement over x402 in USDC and Open USD.
• Our 25 MCP endpoints and 24 JSON APIs are exactly the resource class the Gateway was built to meter — and our x402 discovery endpoints already rehearse the 402 flow (labeled illustrative until enrolled).
• Pilot plan is written and gated: enroll → price one zone sub-cent per call → verify end-to-end → extend fleet-wide.
• The GTM motion is coworker-run: a scored account book — 50 zones ranked weekly for monetization readiness by the sales coworker's own tooling — feeding a gate-owned enrollment pipeline and a weekly forecast line. Adapted from Anthropic's published GTM-engineering and 4,000-account-book practices.

operating model

The org chart is software.

Anthropic defines the primitive: plugins "bundle MCP connectors, skills, slash commands, and sub-agents into a single shareable unit — turning Claude into a specialist tailored to your role" (claude.com/docs/plugins/overview). We instantiated it 20 times — the 12 knowledge-work domains plus deploy, analytics-engineering, agent-resources, and the Apple surface built out into four coworkers (ios-, swift-, hardware-, and xcode-engineering), then a data-engineering coworker to model the fleet's own session telemetry — each with typed skills, least-privilege tools, and explicit approval gates. Spend, outreach, legal sends, and pricing changes always stop at the human.

20 coworkers are registered (subagentidentities.com); 16 are pictured below — the 4 newest (swift-, hardware-, xcode-, and data-engineering coworkers, added 2026-07-04/05) don't have critter art yet, flagged here rather than silently omitted. Fifteenth pictured: agent-resources, onboarded by its own lifecycle skill. Sixteenth pictured: ios-engineering — Swift apps above, Linux VMs below (apple/container), both sides of Apple silicon. Roster: subagentcoworkers.com/platform/plugins

engineering · why one person can run 70 workers

The fleet audits itself.

Marginal operational cost per domain falls as the fleet grows. That's the moat a solo founder actually needs.

unit economics · the complete cost base

Two line items. That's the entire P&L.

• Marginal cost of domain #51 ≈ one $10/yr renewal. Marginal cost of an agent request on Workers ≈ ~0 at current volume.
• No payroll, no SaaS stack, no VMs — the AI workforce cost is flat while the fleet grows.
• Credit honesty: vendor credits are booked at realizable value, never face value — we verified a "$100k credits" headline down to its per-product caps and exclusions before letting it near this plan.

workers launchpad · benefit → what we'd do with it

We're the program thesis, already running.

roadmap & risks · stated plainly

What's next — and what we're not hiding.

Honesty is the house style — our sites publish their own gaps (adoption matrix included). Investors get the same treatment.

the ask

A cohort seat for the network that's already live.

A working agent-readable content network, Cloudflare-native monetization one enrollment away, and a cost base that makes every credit dollar count double.

https://subagentdomains.com              # the 50-domain directory
https://subagentapi.com/api/catalog      # every API, machine-readable
https://subagentdevices.com/architecture # live dependency map
https://subagentcoworkers.com/platform/plugins # the AI workforce

Contact: lexi@opensubagents.com

Prepared 2026-07-02, refreshed 2026-07-14 by the subagent*.com coworker platform (finance, legal, and marketing coworkers — human-gated), inside Cowork/Claude Code — "the same agentic architecture that powers Claude Code," run as a project with standing instructions and durable memory (claude.com/docs/cowork). The company is built inside the surface it documents. Metrics live-verified 2026-07-14 against subagentdomains.com, subagentapi.com/api/catalog, and subagentevaluations.com/api/evaluations — see scripts/flywheel-metrics.sh; Launchpad benefits quoted from cloudflare.com/startups/workers-launchpad. Open USD is announced but not yet on-chain; the Monetization Gateway is early-access — every statement about either is dated and sourced, never projected. Slide-01 artwork: deterministic seeded SVG, generated in-medium by the marketing coworker to creative brief HB-01. Share-card and closing-band artwork: Higgsfield (z_image via MCP + CLI, free tier, briefs OG-01/CL-01, 2026-07-03), palette-retouched and composed by the marketing coworker to the design coworker's briefs; the fifteen pixel coworkers are an original character (brief CH-01) in the genre Anthropic's own Fable 5 announcement critter made delightful — inspiration credited, pixels not traced — all AI imagery provenance-tagged, never passed off as photography. This deck itself meets Claude Science's reproducibility bar: built by a checked-in script (scripts/deck-build.py) from a versioned source, so every slide carries its exact code, environment, and history.