curl GETs of standard public discovery paths across every public Regen host, a soft-404 control for byte comparison, the public GitHub API, and read-only calls to the public regen-compute MCP (the protocol that lets an AI assistant call your tools) server (no wallet configured, no transaction executed). Captured 2026-07-16. This is not a cold audit: the operator is a Regen Tokenomics WG steward with two OPEN public PRs against Regen's own repos, re-verified live at capture. Full reproduction commands: evidence.json.Baseline, credited first, in detail
compute.regen.network ships three parallel machine-readable agent-discovery manifests: an MCP server card (/.well-known/mcp/server-card.json), a Google A2A (a protocol that lets AI agents discover and call each other) agent card (/.well-known/agent.json), and a custom multi-step workflow manifest (/.well-known/agents.json), plus a real llms.txt, a real robots.txt with a Sitemap: line, a complete public OpenAPI spec (13 endpoints), and a standards-conformant self-settling x402 payment flow (exact scheme, USDC on Base, zero signup). That's a stronger three-manifest discovery pattern than any other target we've audited to date, most large companies ship at most one of these three. This is not a "you're behind" brief, it's "the reading-and-paying room is excellent; the field-verification room next door is still empty."The bottom line, in plain English
A retire call that always finishes, or says plainly it didn't → the agent buyer isn't quietly handed back to a human.
Today POST /retire can return success yet retire nothing when there's no wallet or no live liquidity, so an automated buyer's clean end-to-end purchase silently becomes a credit-card link and stalls.
A marketplace agents can actually read → your real credit inventory shows up when an AI goes shopping.
app.regen.network, where the 11 credit classes and 85 projects live (Regen's own live count as of 2026-08-18), returns the same blank app shell to every request, so an agent asking "what can I buy right now" sees nothing and looks elsewhere.
A one-line robots.txt + llms.txt pointer on your main sites → every model represents Regen accurately, at almost no cost.
Your flagship, registry, and docs ship empty robots files, so the excellent agent-legibility one product over never reaches the hosts buyers land on first.
Verified field work underneath the credit → the trust a buyer pays a premium for becomes checkable.
A credit exists because someone did real work in a real place, but nothing funds or proves that visit as a bounded, on-chain-settled receipt, the layer that turns "trust our MRV" into "check the chain."
Finding 1: Regen's own API documents the exact gap
The flagship POST /retire can return HTTP 200 and still have retired nothing.
This is not our inference, it is Regen's own documented behavior, quoted verbatim from compute.regen.network/api/v1/openapi.json: the endpoint can return status:"success" (a real on-chain retirement) or status:"marketplace_link" (no on-chain action at all: no matching sell orders, insufficient supply, a broadcast failure, or a non-zero tx code), and warns downstream systems in writing to check status === "success" before trusting it. Independently, the same MCP server's own retire_credits tool description confirms the no-wallet path is link-only by design, not an edge case: "No crypto wallet needed… Returns a direct marketplace link and step-by-step instructions."
curl -s https://compute.regen.network/api/v1/openapi.json | python3 -m json.tool | grep -A3 marketplace_linkFinding 2: the marketplace itself is invisible to agents
app.regen.network, where the actual credit inventory sits, has zero agent-discovery surface.
The Regen Compute product is agent-ready. The marketplace, the host that actually carries the 11 credit classes / 85 active projects / ~6.1M issued / ~1.4M retired that Regen's own on-chain impact summary reports (live figures as of 2026-08-18), is not. Every one of /llms.txt, /ai.txt, /robots.txt, /.well-known/agent-card.json, /.well-known/ai-plugin.json, /.well-known/security.txt, a real project page (/projects/wilmot), and a nonsense control path returns HTTP 200 with an essentially identical ~2.054 MB client-rendered app shell.
regen-compute MCP server, not through any standard the marketplace itself exposes.curl -s -o /dev/null -w '%{http_code} %{size_download}\n' https://app.regen.network/projects/wilmotFinding 3: the flagship, registry, and docs hosts carry none of the pattern
www.regen.network and www.registry.regen.network each ship a genuine HTTP 200 robots.txt that is literally empty.
www.regen.network: /llms.txt, /ai.txt, /sitemap.xml are true 404s (byte-identical to a control path, a soft-404 SPA shell); /.well-known/{agent-card.json,ai-plugin.json,security.txt} return a different, dedicated 404 (a literal "Invalid .well-known request" page, the app explicitly routes that path, but no card is registered there). /robots.txt is a real 200 that is zero bytes: no Disallow, no Sitemap line, no AI-crawler policy at all. registry.regen.network 301-redirects to a separate Webflow-hosted www.registry.regen.network, whose /llms.txt//ai.txt//sitemap.xml are true 404s (Webflow's default "Not Found" template, byte-identical to a control), and whose /robots.txt is also a real, completely empty 200. docs.regen.network's agent-discovery paths are true 404s too; only its /sitemap.xml is genuine.
Sitemap: line and a pointer to the llms.txt Regen Compute already knows how to write.curl -s https://www.regen.network/robots.txt | wc -cFinding 4: real, useful, undocumented-as-agent surfaces
guides.regen.network ships a real llms.txt; the governance forum ships a real JSON API. Neither is framed as "for agents."
guides.regen.network/llms.txt (200, text/markdown, 13.3 KB, GitBook-generated) enumerates the entire Regen Guidebook, including a real "Buy Ecocredits → Buy with Credit Card / Buy with Crypto → Retire Ecocredits" workflow section. forum.regen.network exposes the full standard Discourse JSON API (/site.json, /latest.json, both real 200s); governance history is already machine-readable at the protocol level. Neither surface is named anywhere as an agent-discovery door; both would take one small doc update to point at.
curl -s https://guides.regen.network/llms.txt | head -5Finding 5: the gap underneath all of it, nobody verifies the field labor
A credit only exists because someone did verified work in a real place. That work has no bounded, funded, proof-verified, on-chain-settled packet, anywhere.
Every finding above is about reading and paying for already-issued credits. None of it touches the step before issuance: an MRV monitoring visit, a restoration planting, a land-steward verification walk. EcoWealth already runs six live x402 programs keyed directly to Regen's own data model: regen-reputation ($0.005), regen-class-momentum ($0.006), regen-retirement-builder ($0.02), regen-methodology-draft ($0.50), regen-rfc-reviewer ($0.10), regen-oracle ($0.005), all live, all at zero paid traffic to date. Useful, but none of the six funds or proves that a human or agent actually walked a hectare, photographed a planting, or GPS-stamped an MRV data point.
curl -s https://vealth.net/.well-known/x402 | grep -A1 regen-The ask
One funded, proof-verified field-work packet, tied to one real Regen project. That's the whole first step.
Fund one EWP work packet: a monitoring visit, a planting-verification walk, an MRV photo set, near a real, already-registered Regen project (e.g. one of the 85 active projects Regen's own impact summary names as of 2026-08-18), paid by any x402-aware agent for a few cents to a few dollars. Deliverable: an on-chain settlement receipt on Base, attachable as supplementary field-verification evidence alongside the credit's existing MRV record.
Or, smaller still: any agent already running npx regen-compute makes one additional keyless x402 call to vealth.net/.well-known/vealth/work-tco2e ($0.005) or /work-packet ($0.10), same standard, same wallet, zero new integration.
Agents start at vealth.net/mcp.
