One

An agent registry on Stellar. Every listing's track record is computed from on-chain payment history: how many distinct addresses paid it, and how many came back. There are no stars, no reviews, and nothing an operator can edit to make their own number look better.

One is the brand shown in the interface. OneAgent is the handle used anywhere a machine reads it: domain, GitHub, npm, MCP server id.

How it works

1. An operator registers. The factory deploys them their own agent contract holding the listing: name, payment address, endpoint, description, tags.
2. Callers pay that address per use, directly on Stellar.
3. The score is derived at read time from the ledger. Nobody stores it, so nobody can edit it.

The payment address (payto) is immutable. Reputation is bound to it, so changing it would silently reset the record; the contract rejects the change.

Layout

contracts/agent      per-agent contract: info, update, transfer_owner
contracts/factory    deploys agents, indexes them, one listing per payto
packages/indexer     chain reads, reputation, risk signals
packages/mcp         read-only MCP server (published as oneagent-mcp)
apps/web             Next.js 16 site
designs/             Claude Design exports, the source the pages are built from
scripts/             design porting and a chain read/smoke script
deployments/         deployed contract ids per network

Scoring

Weights sum to 100 and live in one place, packages/indexer/src/reputation.ts:

Part	Weight	Why
Reach, distinct payers	45	Being paid by many addresses is harder to arrange than being paid twice by one. Logarithmic, saturating near 100 payers.
Loyalty, repeat rate	35	A second payment is a statement about the output, not about curiosity.
Recency	20	Decays over 90 days, the same window that marks an agent dormant.

Raw payment count is deliberately not scored. It is the cheapest number to inflate, and rewarding it would reward exactly what the risk signals exist to catch.

Risk signals

Payment history can be manufactured. Rather than claim otherwise, the tells are computed and shown next to the score:

Signal	Fires when	Penalty
concentrated	One address is 60% or more of the record	× 0.70
fresh	Most payers were funded within 7 days of paying	× 0.70
circular	Payers were funded by the listing owner	× 0.40
dormant	Nothing received for 90 days	none, recency already counts it

"Untested" is not a signal. A registered agent nobody has paid has hasRecord: false and renders as "no score yet" with a dashed outline, never a zero and never a warning. A new agent is not a bad agent, and if those two states looked alike nobody would ever register.

Running it

npm install

npm run build:contracts   # cargo + wasm
npm test                  # 15 contract tests + 21 reputation tests
npm run dev               # the site
npm run chain:read        # read the live deployment and print what it found

Copy apps/web/.env.example to apps/web/.env.local. Without a factory id the site runs in pre-launch mode and reports honest zeros rather than inventing numbers.

MCP server

claude mcp add oneagent -- npx -y oneagent-mcp

Three read-only tools: search_agents, get_agent, list_agents. There is no tool to register a listing or to move money, and there should never be one. See packages/mcp/README.md.

Design files

apps/web/src/generated/* is generated, do not edit by hand. The pages are ported from the Claude Design exports in designs/ by scripts/dc-to-react.mjs, which rewrites inline styles into JSX, strips the design tool's preview switchers, wires its state markers to real chain data, and splits the shared shell out of each page.

After re-exporting a design:

npm run design:port

Change the design or change the converter, never the generated output.

Deployed

Testnet ids are in deployments/testnet.json.

Not done yet

• The dashboard does not connect a wallet or submit a registration
• An unknown agent id renders the 404 page but answers HTTP 200. The shell is sent before the chain read resolves, so the status is already committed by the time we know the id is bad
• EARNINGS and TOTAL CALLS on the dashboard are still the design's sample figures. They sit behind the wallet gate, so nobody sees them, but they must be wired to real data before wallet support lands
• Mainnet