Tokru
An agent-native marketplace for makers, builders, and creators. Stand up a brand, a website, and a storefront that AI agents can actually discover, query, and transact with — built on a typed, A2A-compliant capability profile with full MCP support.
Tokru
Live at tokru.ai
“You own your profile, your AI manages it, Tokru verifies it.”
An agent-native marketplace for makers, builders, and creators — a way to stand up a brand, a website, and a storefront that AI agents can actually discover, query, and transact with.
The Problem
If you’re an independent maker, freelancer, or small storefront, the thing sizing you up first is increasingly not a person. It’s an agent, working through a shortlist on someone’s behalf.
Everything we currently hand that agent is wrong for it. A portfolio site is a human artifact — layout, copy, images, implied structure. The agent scrapes it, guesses at what’s a skill versus a section header, misses your availability entirely, and has no way to verify that any of it is true. You lose to whoever happened to be more legible.
Tokru inverts the order of operations. The machine-readable profile is the source of truth. The human-facing site is rendered from it. Both stay in sync because neither is a copy of the other.
Every Profile Is Also an Agent
A Tokru profile publishes at {handle}.tokru.ai and is two things at the same address:
- a website a person can read, and
- an agent endpoint, serving its own
/.well-known/agent-card.json, its own/.well-known/tokru.jsoncapability profile, and its own A2A JSON-RPC endpoint at/a2a.
The registry at tokru.ai is itself an A2A agent. It advertises a capability-search skill: an agent queries by skill, availability, or domain and gets back A2A-compatible agent cards for the profiles that match. Discovery happens agent-to-agent — no SEO, no crawling, no guessing. For authenticated agent access there’s an MCP server at mcp.tokru.ai.
You can see the whole thing working on my own profile: abstractryan.tokru.ai.
Nobody Should Start From a Blank Page
The hard part of any builder isn’t rendering. It’s the empty form.
Tokru’s wedge is rebuild, don’t author. Point it at a GitHub account, an existing portfolio URL, or a resume. It scrapes or parses what’s there, an LLM extracts it into the typed profile schema, and you’re handed a pre-filled draft to review and correct — not a blank canvas to fill.
The detail that makes this work is that the extraction target is a typed schema, not freeform text. Most AI site builders generate prose and markup directly, which is why their output is structurally interchangeable and largely illegible to machines. Extracting into a contract first means the result is machine-readable by construction. The typed profile isn’t a feature of the product — it’s the whole moat.
One Pipeline, Three Tracks
The same engine drives three tracks — website, portfolio, and storefront. What changes per track is the connectors, the output template, and which parts of the schema matter most. What doesn’t change is the extraction-to-typed-profile core.
The portfolio track’s reference implementation is abstractRyan.ai — this site — vendored into the Tokru repo as a working quarry. The storefront track’s is StoreFrontBuilder. Ideas get proven on a real site under real load before they become something a user can push a button on.
Technical Architecture
- Cloudflare Workers + D1 — a single registry Worker serving
tokru.aiand every{handle}.tokru.aiprofile from the edge - Clerk — authentication and session management
- Cloudflare Workflows — the onboarding build pipeline, with drafts persisted in D1 so a rebuild can be resumed or reviewed rather than restarted
- Capability profile schema (v1/v2) — the typed contract every surface renders from
- A2A + MCP — the registry, each profile, and the MCP server, all first-class
- Admin API — account management with full audit logging on every mutation; no direct writes to production data
create-tokru-profileCLI — publish a machine-readable capability profile from the terminal in under fifteen minutes
Deploys apply D1 migrations ahead of the code push, so schema always ships with the release that needs it.
Status
Live, in production, and in active development. The current focus is the portfolio track of AI onboarding — rebuilding a user’s professional presence from a GitHub account, URL, or resume into a full Tokru profile and a richer {handle}.tokru.ai site.
Tokru is where the agent-discovery experiments from abstractRyan.ai get generalized into something other people can actually use.