diff --git a/skills/agentflow/SKILL.md b/skills/agentflow/SKILL.md
new file mode 100644
index 0000000..d025657
--- /dev/null
+++ b/skills/agentflow/SKILL.md
@@ -0,0 +1,89 @@
+---
+title: "AgentFlow — wallet trust & sybil screening on Base"
+description: "Before an agent transacts with, trades against, pays, or trusts an unknown Base wallet, screen it for wash-trading / sybil behavior. Returns a calibrated sybil score (0–100) + confidence + the reasons, from real on-chain USDC flow. The safety layer for agentic payments — pairs with the transact/trade/lend/buy skills."
+name: agentflow
+version: 0.1.0
+---
+
+# AgentFlow — is this Base wallet a real agent, or a sybil?
+
+The onchain agent kit lets an agent **transact, trade, lend, mint, and buy**. AgentFlow is the
+step that should come *first*: **screen the counterparty.** Before sending USDC to, trading with,
+or trusting an unknown Base address, ask AgentFlow whether it's a genuine autonomous agent or a
+wash/sybil wallet inflating its own volume to look legit.
+
+Behavioral heuristic on public Base USDC flow — **not financial advice**, and a probability, not a
+verdict. The caller owns the threshold.
+
+## Detection
+
+If AgentFlow MCP tools (e.g. `screen_wallet`, `verify_agent`, `get_counterparty_risk`) are
+available in the harness, use them directly. Otherwise call the HTTPS endpoints below — no SDK,
+no signup. The free demo endpoint needs nothing; the paid endpoints settle in USDC via x402
+(the agent pays directly, no account).
+
+## When to use it
+
+Trigger a screen **before** any of these, when the counterparty is an unknown Base address:
+- About to send/transfer USDC to a wallet
+- About to trade against, LP into, or route through an address
+- About to trust a "counterparty agent" claiming to be autonomous
+- Evaluating whether inbound funds came from a real agent or a wash farm
+
+Skip it for well-known infrastructure (DEX routers, CEX wallets, bridges) — AgentFlow already
+labels those as hubs, not agents.
+
+## How to call it
+
+- **Free (rate-limited, great for one-off checks):**
+  `GET https://agentflow.watch/public/wash-demo?address=0x...`
+- **Paid, unlimited, machine-native ($0.01/call via x402):**
+  `GET https://agentflow.watch/v1/wash?address=0x...`
+- **Identity card — "is this a real agent, and who?" ($0.01):**
+  `GET https://agentflow.watch/v1/agent/0x...`
+- **Deep counterparty risk verdict ($0.05):**
+  `GET https://agentflow.watch/v1/risk?address=0x...`
+- **Signed, off-platform-verifiable receipt (EIP-712, $0.10):**
+  `GET https://agentflow.watch/v1/attest?address=0x...`
+- **Free benchmark context:** `GET /index.json` (real-vs-wash), `GET /pulse.json` (agents active now)
+
+For volume: buy a bulk pack (`/v1/credits/buy`, $5 / 1,000) or a 30-day unlimited pass
+(`/v1/subscribe`, $19), then pass the returned key as the `X-AgentFlow-Key` header.
+
+## How to read the result
+
+```json
+{
+  "sybil_score": 81,            // 0–100, higher = more wash/sybil-suspect; null = not enough history
+  "confidence": 0.9,            // 0–1
+  "confidence_level": "HIGH",   // LOW | MEDIUM | HIGH (scales with sample depth)
+  "recommendation": "high wash/sybil signal — screen further or avoid",
+  "reasons": ["round-trip ratio 83%: received USDC returned to sender within 20% / 1h", "..."],
+  "signal": { "wash_ratio": 0.83, "is_hub": false, "agent_score": 70, "known_entity": null, "active": true }
+}
+```
+
+Suggested handling (the caller sets the final threshold):
+- **`sybil_score` null** → insufficient on-chain history; don't treat "unknown" as "safe." Require more
+  history or proceed with caution.
+- **≥ 60** → high wash/sybil signal. Avoid, or require additional proof, before sending funds.
+- **50–59** → elevated; proceed with caution.
+- **25–49** → some round-trip activity; minor caution.
+- **< 25** → clean on the round-trip heuristic.
+- Always weigh `confidence_level` — a HIGH-score / LOW-confidence result means "thin sample, verify."
+
+## Method (so you can trust it)
+
+Wash/sybil = round-trip ratio: received USDC that returns to the sender within ~20% amount and ~1h,
+among hub-filtered wallets (infrastructure stripped out). Score is a calibrated probability;
+confidence scales with how many transfers it's built on. Base USDC only. Verify any wallet on
+basescan.org.
+
+## Installation
+
+```
+npx skills add leonidahqadmin-cell/agentflow --skill agentflow
+```
+
+Once installed, the agent uses it automatically when a task involves trusting or transacting with
+an unknown Base wallet. Live API + docs: https://agentflow.watch