AT Directory

Where agents discover OP-verified merchants on the rails Bitcoin and Tether actually use.

AT Directory indexes merchants that sell products, services, APIs, or content and accept Lightning, BOLT12, L402, or USDT (any chain), with trust attestations issued through Observer Protocol. It is the verified, agent-callable layer on top of the fragmented agent-payment landscape. It is not in the payment path — agents discover here and pay merchants directly.

• MCP server (primary product) — @agenticterminal/mcp-server, hosted at mcp.agenticterminal.ai.
• Web directory — agenticterminal.ai, same data rendered for humans.
• SKILL.md — teaches an agent how to query and complete payment handoff.

Status

Pre-v1, active build. Target: week of 2026-05-18 for a live Tether-meeting demo. v1 ships Tiers 1–2.

Use it from an agent

// Hosted (recommended, no install)
{ "mcpServers": { "at-directory": { "url": "https://mcp.agenticterminal.ai/mcp" } } }

npm install -g @agenticterminal/mcp-server   # local alternative

Six tools: search_merchants, get_merchant, verify_payment_endpoint, list_categories, list_rails, whoami. Reads are ungated — anonymous agents see all tiers and can transact with no setup. An Observer Protocol DirectoryAccessCredential raises rate limits and unlocks write access (reviews), but is not required to discover or transact. Full reference: agenticterminal.ai/api-docs.

Repo layout

data/merchants/         One JSON file per merchant (the source of truth).
data/schema/            JSON Schema (Draft 2020-12) for merchant records.
data/categories.json    Category taxonomy.
data/rails.json         Supported rails + USDT chains.
data/LOGO-AUDIT.md      Logo sourcing checklist.
packages/core/          Types, schema-validated load, search/filter.
packages/mcp-server/    stdio + hosted HTTP MCP server, rail verification.
packages/skill/         SKILL.md (canonical).
apps/web/               Next.js static directory + Netlify Functions.
deploy/mcp-server/      Dockerfile, systemd unit, DEPLOY runbook.
scripts/                Data validation (and one-time crawl tooling).

Development

pnpm install
pnpm -r test          # core + mcp-server unit/integration tests
pnpm -r typecheck
pnpm validate-data    # schema-validate every merchant record
pnpm -F @at-directory/web dev          # web at localhost:3000
pnpm -F @agenticterminal/mcp-server start:stdio   # MCP over stdio

Local MCP with the mock verifier (until the real AT route lands):

pnpm -F @agenticterminal/mcp-server exec tsx src/dev/mock-verifier.ts &
AT_VERIFIER_URL=http://127.0.0.1:8787 \
  pnpm -F @agenticterminal/mcp-server start:stdio

Data model

Every merchant carries two orthogonal signals — read both:

• OP trust tier (the merchant): 1 self-attested · 2 operator-verified. v1 Tier 2 = AT operator transacted and confirmed; the bilateral counterparty-attestation model lands in v1.x.
• Agent-callable tier (the integration): full-api · structured-handoff · human-checkout.

The trust tier gates nothing. Reads are ungated, every tier is visible to anonymous agents, and an agent can transact with a Tier 1 merchant with no setup. The tier is evidence for a caller to weigh, not a decision this directory has made on its behalf. Nothing is refused on the basis of it.

Schema is enforced in CI (pnpm validate-data) and on every load. USDT rails require a chain; Tier 2+ non-integrated records require an op_attestation_url; an op_trust_tier of 3 is rejected on ingest, which is why the model above stops at 2.

Contributing

See CONTRIBUTING.md. Merchants can self-register at agenticterminal.ai/submit — submissions open a PR for review and land at Tier 1 until verified.

Documents

• Scope — positioning, inclusion criteria, definition of done.
• Spec — implementation contract.

License

MIT.