Build an AI Agent That Watches Crypto Prices and Pings You

Most crypto price alerts are dumb. They fire when a number crosses a line and dump raw data on you: "BTC: $67,420. +3.2%." You still have to decide whether to care.

This guide builds something different: an always-on AI agent that checks market data on a schedule, reasons about what it sees against your watchlist and thresholds, and sends you a short, plain-English Telegram message that tells you why it's pinging you — not just that it is.

The whole thing runs on AgentRoost's Hermes framework, which combines scheduling, live data tools (MCP), persistent context, and an auto-provisioned Telegram bot. AI credits are included in the subscription, so the reasoning step costs you nothing extra — there's no OpenAI or Anthropic key to wire up.

---

What the finished agent does

Every 30 minutes (or whatever interval you set), your Hermes agent:

1. Pulls current price, 24h change, and volume data for your tracked symbols using a live market MCP tool.
2. Checks each symbol against thresholds you configured at startup (e.g., "alert me if BTC drops below $60k, or if ETH moves more than 5% in an hour").
3. Passes the raw data plus your thresholds to the LLM layer, which reasons about whether the move is worth your attention and what context is relevant.
4. If it decides yes, it sends a concise Telegram DM — in natural language, with a brief rationale.
5. Remembers what it already told you (persistent context), so it doesn't fire the same alert five times in a row.

The key difference from a webhook-based price alert: step 3. The model can say "BTC crossed $60k but volume is thin and it's already bouncing — this looks like a wick, not a breakdown" instead of just "ALERT: BTC below $60k."

---

What you need before you start

• A Telegram account (for receiving alerts).
• A list of symbols you want to watch and your threshold rules — have these ready as plain text.
• An AgentRoost account. Sign up here — $19.99/mo, AI credits included, 14-day money-back guarantee.

That's it. No exchange API keys, no cloud provider, no Docker.

---

Step 1 — Launch your Hermes agent

1. Sign up at AgentRoost and pick the Hermes framework.
2. Name your agent (e.g., market-watcher). It provisions in under two minutes.
3. Open the AgentRoost manager bot on Telegram and /start your new agent. You'll get a private bot link — tap it to open your agent's dedicated Telegram thread.

At this point you have a live, persistent AI assistant with an always-on runtime and a Telegram interface.

---

Step 2 — Define your watchlist and thresholds

In your Hermes agent's first message, give it your rules in plain language. Be specific — the LLM stores this in its persistent context and references it every check cycle:

Watch these symbols: BTC, ETH, SOL, WIF.

Alert rules:
- BTC below $60,000 OR above $72,000
- ETH 1h move > 4% in either direction
- SOL daily volume drops below 30-day average (flag as "unusual low volume")
- WIF: alert on any >8% daily move

Don't alert me more than once per rule per 2-hour window unless the price
moves another 3% from the last alert level.

Hermes stores this as its operating instructions for the session. Because context is persistent across restarts and days, you only do this once (or update it when your rules change).

---

Step 3 — Enable the live market data tool

Hermes supports MCP (Model Context Protocol) tool servers, which give the LLM access to live external data. The market-data MCP provides real-time price quotes, 24h stats, and volume data for major crypto symbols.

Tell your agent to use it:

Use the market data tool to check prices on a schedule.
Start checking every 30 minutes, beginning now.

The agent will confirm the schedule and tool binding. Under the hood, it's wiring the MCP tool into the scheduled task loop — you don't configure anything manually.

If you want finer control over the schedule, you can specify cron syntax directly: "Check at minutes 0 and 30 of every hour" works just as well.

---

Step 4 — What the reasoning step looks like

When the scheduled check runs, Hermes calls the market MCP tool to get current data, then reasons against your rules. A typical internal pass looks roughly like this:

[Tool result]
BTC: $59,840 (-2.1% / 1h), volume: normal
ETH: $3,210 (-1.2% / 1h), volume: normal
SOL: $142 (-0.8% / 1h), volume: 22% below 30d avg ← LOW
WIF: $2.41 (-1.1% / 1h), volume: normal

[Rule check]
BTC: $59,840 < $60,000 threshold → triggered
SOL: volume flag triggered

[Context check]
Last BTC alert was at $60,100, 55 minutes ago — 2-hour window active → SKIP
SOL volume flag: not fired in last 2 hours → SEND

You receive:

SOL — unusual low volume SOL is trading at $142 with volume 22% below its 30-day average. Price is only down ~0.8% on the hour, so this isn't a panic sell-off — it's quiet. Could be a weekend lull or could precede a sharp move either way. Worth watching. [05:30 UTC]

That's the reasoning output: not a raw number, but a short interpretation with relevant context. The de-duplication logic (the 2-hour window rule) comes from the persistent instructions you set in Step 2.

---

Step 5 — Tune it after the first day

After running for 24 hours, you'll have a sense of whether alerts are too noisy or too quiet. You can update your rules mid-conversation:

• "Tighten the ETH threshold to 6% moves."
• "Add AVAX to the watchlist with the same rules as SOL."
• "Don't alert me between 00:00 and 07:00 UTC."

Hermes updates its context immediately and applies the new rules on the next cycle. No YAML to edit, no config file to reload.

---

Tips and pitfalls

Symbol naming. Use the exact ticker symbols the MCP tool supports (BTC, ETH, SOL — not "Bitcoin" or "bitcoin"). If a symbol isn't found, the tool returns an error and the agent will tell you.

Don't over-specify thresholds at launch. Start with 2-3 symbols and your most important rules, then add more after you've seen the alert cadence live. Overly tight thresholds on a volatile market will produce alert fatigue fast.

The de-duplication rule is the most important one. Without it, a sideways chop around your threshold will ping you every 30 minutes. The "don't re-alert within 2 hours unless price moves another X%" pattern is worth writing explicitly into your initial instructions.

Schedule frequency vs. model calls. Every check cycle makes an LLM call to do the reasoning step. More frequent checks mean more model usage. AI credits are included in your subscription, but if you're watching 20 symbols every 5 minutes, you'll burn through them faster. 15-30 minute intervals are a good starting point for most watchlists.

---

Running this on AgentRoost

The entire setup above runs on Hermes, AgentRoost's persistent agent framework. What that gives you:

• Always-on runtime — no Lambda cold starts, no "agent fell asleep," no cronjob to babysit on a VPS.
• AI credits included — the reasoning step in every check cycle uses the built-in LLM budget. No separate OpenAI/Anthropic key, no surprise bills.
• Telegram bot auto-provisioned — you get a private, dedicated bot. No BotFather steps, no token to manage.
• Persistent context — your watchlist, threshold rules, and alert history survive restarts and are available to the model on every cycle.
• 350+ LLM models available — if you want a faster model for high-frequency checks or a stronger one for more nuanced reasoning, switch anytime from the dashboard.

Plans start at $19.99/mo, cancel anytime. Get started with Hermes or compare plans.

---

What this isn't

This agent interprets market data and notifies you — it doesn't execute trades, connect to your exchange account, or manage a portfolio. It's an observation-and-reasoning layer that replaces "I need to stare at charts all day" with "ping me when something worth looking at happens, and tell me why."