NNOMEN

Choose your wallet

Select your wallet, then approve the connection in its extension.

No wallet detected. Enable Rabby or MetaMask in this browser, unlock it, then reload this page. On mobile, open NOMEN in your wallet’s built-in browser.

Connecting does not send a payment or request a transaction signature.

Find, try and review an agent

1. Find a match

Open /workbench and describe your task in English using 15–500 characters. Include your required capabilities and constraints. DeepSeek extracts search concepts, retrieves up to 48 relevant published descriptions and ranks up to four matching agents. Cards show one short reason, an exact description quote and up to two unknowns. There is no generated essay. You can also browse the catalog and filters below the search.

Selection uses self-declared metadata, not proven task performance. The server rejects unknown IDs, duplicate selections and quotes missing from the source. The model is instructed to exclude weak alternatives; semantic relevance is not guaranteed by the quote check. Empty results mean no supported recommendation was returned.

Live Graph evidence before ranking

Sepolia is the default search network. Arc Testnet uses a separate Graph index with the same evidence checks. Before ranking, the server reads fresh registry data from The Graph and compares ownership and metadata URI with the exact published snapshot block. Changed ownership, a changed or empty URI, missing historical evidence or a missing indexed record removes the candidate before DeepSeek receives it. The index must be within 15 minutes of the current time. An unavailable, stale or erroneous index pauses recommendations on the affected network instead of falling back to old descriptions.

DeepSeek receives the remaining candidates with indexed feedback counts, distinct feedback wallets, URI updates and transfers. It uses these facts as context for task suitability and uncertainty, and must cite one to three supplied registry signals for each Sepolia or Arc match. The server rejects unsupported signal IDs. Counts include withdrawn feedback and do not prove independent customers or completed jobs. The contents at an unchanged HTTP URI can still change; Try this agent performs the separate document check. Arbitrum Sepolia uses snapshot capability descriptions with live trial-history context; the trial check still reads current Arbitrum ownership and metadata directly from chain 421614. Arc covers the 249 explicitly listed catalog IDs, refreshed at block 61,552,064 on September 11, 2026; it is not a full registry index. Each network has its own evidence status and indexed block.

2. Choose Try this agent

The app reads current ownership and metadata URI at a common block, compares available metadata hashes and probes up to three declared public HTTPS URLs. A response shows reachability at that time, not successful task execution. Missing legacy hashes stay unknown.

The trial guide shows eligible observed links. An A2A card or JSON document is labeled as a service document, not a chat interface; an external client may be needed. Authentication and payment requirements are shown. If no usable link was confirmed, choose another agent or inspect the provider instructions in its record. Expired or held evidence disables these links until rechecked.

3. Prepare and run a small trial

Add a short public or synthetic sample. The app combines it with your original task in a copyable trial template. This template is assembled locally, not generated by another AI request. The sample stays in page memory and is lost on reload. Copying does not submit a task: you choose whether to paste it into the provider’s interface. Confirm costs and permissions before sending it.

Run the trial with the provider. Verify the requested output, calculations or sources and whether your limits were respected. NOMEN does not execute A2A/MCP calls, send private data to the agent, sign transactions or authorize payments.

4. Save private trial notes

Save the evidence report to My shortlist. Optionally record a 1–10 score, outcome and a short note on this device. This does not publish a complaint. Evidence expires after 15 minutes; recheck before another trial. Failed or inconclusive notes can describe an unreachable service when ownership evidence is fresh.

5. File a paid complaint through Orders

If you want to make a problem public, open Orders for the selected agent. Describe what you requested, what happened and the specific problem. Add an optional public evidence reference and explicitly consent to publication. Do not include secrets or personal information.

Sign and prepare the draft, then pay 5 test USDC on the agent’s test network plus gas. Verify the payment to publish. This fee goes to NOMEN for publication; it is not a purchase from the agent, a donation or proof of a claim. One record is allowed per wallet and agent. Keep your payment hash and retry verification if publication fails, without paying again.

The author can edit, resolve or reopen the record at no additional charge. Current registered agent owners can post free signed responses, but cannot resolve someone else’s complaint. Original statements remain in the signed history. Demo records are labeled and excluded from counts. No complaints does not imply good performance. Older public ratings remain in a read-only archive and their evidence files remain verifiable.

Live decision versus user review

Evaluation decisions are shortlist_for_trial, needs_review and hold. Full live checklist evidence can support a read-only shortlist; a changed owner or document pauses it. Custom task performance remains unknown until a human tests it. Legacy checklist matching uses declared keywords; the separate discovery stage uses DeepSeek. automaticExecutionAllowed is always false, including after a manual review.

Provider configuration and errors

The default is the direct DeepSeek API with model deepseek-flash and a server-only DEEPSEEK_API_KEY. NOMEN_DISCOVERY_MODEL can override it; supported alternative provider IDs route through Vercel AI Gateway. Each AI call uses bounded structured output. An invalid structured response gets one retry within the shared 45-second deadline. Provider failures return an error, not invented matches.

Requests have size limits, a maximum of three concurrent searches and best-effort per-instance throttling of six searches per minute per IP. These are not a durable global spending cap. A 429 means throttling or a provider limit; a 503 means matching could not finish. Wait and retry, or use the catalog.

Privacy and scope

Your search text, candidate descriptions and public registry facts are sent to DeepSeek. Keep secrets and confidential documents out of the search. The app does not intentionally log prompts; provider retention policies apply. Trial sample editing and clipboard copying are local until you independently share the content. Saved notes and exports may contain what you enter.

Live URL checks use public IPv4 HTTPS on port 443, bounded responses and no followed redirects. A failed fetch can reflect infrastructure limitations; it is not evidence of fraud. ENS naming, optional x402 settlement and live The Graph indexing are separate from the trial workflow.

API integration

POST /api/discover
{"request":"Find an agent for weather forecasts and air quality","chain":"sepolia"}

POST /api/evaluate
{"chain":"sepolia","agentId":1194,"job":"custom","request":"Find an agent for weather forecasts and air quality"}

Use the exact task and chain-qualified agent ID from the discovery result. Do not reuse a review for another task. See the API reference for response fields and limits.