Automate Your Crypto Reports With an AI Agent on Telegram

Most Crypto Telegram Bots Want Your Keys. You Probably Don't Need One.

Open any "best Telegram crypto bots" list from the last six months and you will see the same names: GMGN, BullX, Trojan, BONKbot, 3Commas. They all do the same dance. You connect a wallet, hand over signing permissions, and the bot trades on your behalf. The tradeoffs are well documented now, and the news in 2025 and 2026 has not been kind:

A crypto trader was drained of $200,000 in May 2026 through a Telegram-bot-linked compromise.
Security researchers identified over 340 fraudulent crypto trading bots on Telegram during 2025, with an average lifespan of 45 days before being reported and removed.
In Q3 2025 alone, Solana Telegram bots were linked to more than $889 million in fraud, including the BONKbot and Solareum exploits.

If you opened the same lists looking for "I want to know what happened to my bag overnight without checking five tabs," you walked away with a list of trading bots you do not need and a security model you cannot fully trust. The thing you actually want is a daily report, not a trading agent. Those are completely different problems and the second one is fully solvable without giving anyone your keys.

This post shows you how to build that daily report with a non-custodial AI agent on Telegram, using only read-only APIs and a few small scheduled tasks.

What a Useful Crypto Report Actually Contains

Before you automate anything, write the report you wish you woke up to. For most self-custody crypto holders the useful set is small:

Portfolio delta. Total value now vs 24 hours ago, broken down by ticker. One line per position.
Tracked wallet moves. A small list of wallets you care about (a whale, a project treasury, your own cold storage) with their net inflow or outflow overnight.
News on your tickers. Three to five headlines from sources you trust, filtered to coins you actually hold.
Gas snapshot. If you are EVM-active, the current base fee in gwei and the trend over the last day.
One contextual line. "BTC is up 2.3% but underperforming ETH by 1.4 points" hits differently than a list of numbers.

That is a 10 line message. You do not need a $99 per month dashboard. You need three small scheduled tasks and a Telegram chat.

The Recipe: Three Chained Jobs

The trick is splitting the work across three jobs so the expensive part - the AI step - only runs once a day. The other two are cheap polls.

Job 1: Cheap price poll (no LLM)

A small script runs every 30 minutes. It hits the CoinGecko public API (free tier supports 30 requests per minute with a demo account) to fetch current prices for your tickers, calls your wallet RPC for balances, and writes the result to a local file. No AI is involved.

In Hermes Agent terms this is a no_agent=True cron job. The agent does not wake up at all. Your script runs, the output is saved, and the scheduler moves on. This pattern is what makes the whole thing affordable to run on a 30 minute cadence.

hermes cron create "every 30m" --no-agent \
  --script crypto/fetch_prices.sh \
  --name "Price poll"

Job 2: Hourly news sweep

A second job runs every hour using the blogwatcher skill. It reads a small list of RSS feeds (CoinDesk, The Block, plus any niche feeds you follow), filters to articles that mention your tickers, and writes a short list of headline + URL pairs.

hermes cron create "every 1h" \
  --skill blogwatcher \
  --prompt "Scan the configured feeds for articles about ETH, SOL, ARB. \
            Return a deduplicated list of headline + URL." \
  --name "Crypto news sweep"

Job 3: The 8am digest (the only agent-heavy step)

This is the job where the agent earns its keep. It uses Hermes's context_from feature to prepend the output of the price job and the news job as context, then writes the actual report in your voice.

hermes cron create "0 8 * * *" \
  --context-from "Price poll,Crypto news sweep" \
  --prompt "You are my morning crypto desk. Using the price data and \
            news context above, write a 10 line Telegram message: \
            (1) portfolio total + 24h delta, \
            (2) per-position delta sorted by absolute change, \
            (3) any tracked-wallet moves above 50 ETH equivalent, \
            (4) the 3 most material headlines, \
            (5) one contextual paragraph in my voice." \
  --name "Morning crypto digest"

At 8am, the agent wakes up fresh, sees the price snapshot and the news list as context, composes the digest, and delivers it back to whatever Telegram chat created the job. The whole thing takes seconds. You read it before your coffee gets cold.

The natural-language version of that schedule works too. "Every day at 8am, summarize last night's portfolio changes and the news on my tickers, then send it to me on Telegram" gets parsed into the same job. The Hermes scheduler accepts either.

Why "Non-Custodial" Is the Whole Point

The reason this pattern works is that the agent never holds anything custodial. The wallets it reads are read-only RPC calls. The exchange API keys, if you use any, are restricted to read scope. The LLM key is yours (BYOK), so the prompts and the responses never leave your trust boundary. There is no "agent wallet" sitting somewhere with permission to sign.

This is the same architectural insight that pushed Bitbot to position itself as "the world's first self-custodial Telegram trading bot" and that pushed TON Tech to publish an Agentic Wallets standard in April 2026 for non-custodial agent spending. The market is finally catching on. You can use the same idea today for the simpler problem of just reading and reporting.

The honest tradeoff: you give up automated execution. Hermes will not place a market order for you on a price trigger. That is intentional. If you want automated execution, you accept the custodial risk that comes with it. If you mainly want to know what happened overnight and react like a human in the morning, the read-only approach is the right one.

Adapt the Same Skeleton to Anything

The structure (cheap poll, periodic enrichment, daily agent summary) is not specific to crypto. The same three-job pattern works for:

Stock portfolios. Swap CoinGecko for Alpaca or Yahoo Finance.
NFT floor watching. Reservoir or OpenSea instead of CoinGecko.
DeFi position health. Zerion or DefiLlama for your LP and yield positions.
Stablecoin yield comparison. A daily "best risk-adjusted yield right now" digest from a few protocols you trust.

Once you have one of these recipes working, the others are an afternoon each. The agent does not care what shape the data takes, only that you defined what "useful" means in the prompt.

Where This Runs Matters More Than People Think

Scheduled tasks only fire if your agent is running. A cron job on a laptop that sleeps at midnight is not a cron job, it is a wish. Two reasonable paths:

Self-host on a small VPS. A cheap VPS is enough for one Hermes agent with a handful of cron jobs. The blogwatcher and price poll are tiny workloads. Pair this with the Hermes Agent Docker setup if you want a clean container model.
Use managed hosting. Hermify Starter at $19 per month is the cheapest way to get a Hermes agent running on persistent infrastructure with BYOK for your LLM, your crypto API keys, and your prompts. The same architecture, none of the babysitting.

The full feature set of Hermes scheduled tasks (schedule syntax, lifecycle, delivery targets, skill attachment) is covered in our scheduled tasks pillar post. Read that next if you want the broader picture before you build.

Get Started

If you already have Hermes running, try the price poll first. It is the cheapest of the three jobs and gives you immediate feedback that the chain works. Once you have prices flowing into a file, add the news sweep, then add the 8am digest.

If you do not have Hermes running yet and you want the morning report without spending a weekend setting up a VPS, get started with Hermify and have your agent live in under 60 seconds. The Starter tier is BYOK, which means you keep full ownership of every credential the agent touches. That is the model this whole recipe is built on.