Skip to main content
Glama
bytx-trader

bytx-trader

Official
by byttrade

bytx-skills — crypto trading skills for AI assistants

English | Русский

Eight steps for one crypto trade, from "what's moving" to the order: scan the perps, read funding and open interest, time the entry, plan the exit, replay the rest of the session 10,000 times, size the position from the money you are ready to lose, check it against your limits, and send it — paper by default, never without your yes.

Works with any AI assistant that supports MCP — for example, Claude. The logic is a Python package with an MCP server and a CLI on the bytx exchange library; for Claude there are also skills (/crypto-scanner …) on top.

Paper trading runs on the exchange's live order book with virtual money: no keys, no signup.

Install

You need uv (brew install uv or curl -LsSf https://astral.sh/uv/install.sh | sh).

Any MCP client (Claude Desktop, Claude Code, Cursor, …): add the server to the client's MCP config — templates in mcp-configs/:

{ "mcpServers": { "bytx-trader": {
    "command": "uvx",
    "args": ["--refresh-package", "bytx", "--refresh-package", "bytx-trader",
             "--from", "bytx-trader@latest", "bytx-trader-mcp"],
    "env": { "BYTX_TRADER_HOME": "/absolute/path/to/your/project/crypto" } } } }

Then ask "scan the crypto market" or start from the server's trade_pipeline prompt. Set the confirm tool to ask you every time. Without uv: pip install bytx-trader and "command": "bytx-trader-mcp".

Claude only, on top of that:

/plugin marketplace add byttrade/bytx-skills
/plugin install bytx-trader@bytx-skills

or copy the folders: cp -r bytx-skills/crypto-* ~/.claude/skills/ (or your project's .claude/skills/).

any MCP client

Claude only

all eight steps, paper and demo

MCP tools

risk gate and one-time confirm token

in the server

the trade as a script

trade_pipeline prompt

skills

/crypto-* commands, triggers by phrase, plugin install

✓

Related MCP server: hyperliquid-mcp

The steps

step

MCP tool / CLI

what it does

crypto-scanner

scan

Top perps by 24h volume: breakout of the session's opening range, RSI extremes.

crypto-derivs

derivs

Funding now and 7-day average, open interest change, perp/spot basis, crowded-side flags.

crypto-entry

entry_check

Waiting / triggered / expired, on closed bars only.

crypto-exit

exit_plan

Stop (range-mid, range, ATR, price), target in R, flat-by time.

crypto-timemachine

timemachine

10,000 replays of the rest of the session from the coin's own bars + the same rule on every past session.

crypto-size

size

Size from risk with fees, margin, liquidation estimate, funding cost. Places nothing.

crypto-risk

risk_check

PASS / BLOCK against your limits.

crypto-broker

execute, confirm, flatten, positions, reconcile

The only step that sends anything.

Crypto trades 24/7, so the "session" is yours: crypto/config.json → session.start (00:00 UTC by default), range_minutes, entry_until, flat_by. Everything lands in crypto/ and each step reads what the last one wrote.

An example: BTC on Binance, November 18 2026, real candles

  1. scan at 01:05 UTC: BTC broke its first-hour range (64,673.5–64,976.3) to the downside; the 01:00 bar closed at 64,588.5, 0.131 % below.

  2. entry_check: triggered, short.

  3. exit_plan --stop range-mid --target-r 2: stop 64,824.9 (0.366 % away), target 64,115.7, flat by 23:55 UTC.

  4. timemachine: 10,000 replays from the last 30 sessions (seed 19): target first 37.6 %, stop first 62.4 %. The same rule on those 30 sessions: fired 30 times, 9 targets, 21 stops. Expectancy before fees +0.13R simulated, −0.10R historical. Verdict: close to a coin flip — that is the point of looking before you risk a dollar.

  5. size with 10 USDT at risk, 5x: ≈0.033 BTC, ≈2,131 USDT notional, margin ≈426, loss at the stop ≈9.94 USDT with fees.

  6. risk_check: PASS.

  7. execute shows the ticket and returns a one-time token; you answer yes; confirm sends it to the paper account.

tests/ reproduce these numbers offline from the candles in tests/fixtures/.

Modes and the gates

mode

how

keys

paper (default)

live order book, virtual balance account_usdt, real taker fees

none

demo

"mode": "demo": the exchange's demo/testnet

BYTX_API_KEY, BYTX_SECRET (BYTX_PASSWORD for OKX, KuCoin, Bitget; BYTX_WALLET_ADDRESS, BYTX_PRIVATE_KEY for Hyperliquid) in the server's environment

real

not sent by this package: real orders go through the hosted BYTtrade MCP

kept by BYTtrade

Every order, in every mode, passes three gates, enforced in the code:

  1. the risk check passed this exact ticket today (fingerprint);

  2. execute sends nothing — it returns the ticket and a one-time token (5 min); confirm takes the token and your reply word for word, and anything but a plain yes cancels; the risk check runs again;

  3. your client's own approval prompt for confirm.

The client order id comes from the ticket, so if the exchange times out the order is found by that id with reconcile — it is never sent twice. Getting out (flatten) is never blocked by daily limits.

Updates

Started with uvx as above, the server picks up the newest bytx-trader and the newest compatible bytx every time your assistant starts it: fixes and new exchange features arrive with no action on your side. bytx-trader accepts bytx within one series (>=0.2.1,<0.3), so a breaking bytx release never arrives on its own: it comes with a bytx-trader release that was tested against it. Installed with pip, the server checks PyPI once a day and tells you in its first answer when an update is out (BYTX_TRADER_NO_UPDATE_CHECK=1 turns that off). Claude skills installed as a plugin update with the marketplace (/plugin).

Development

pip install -e ".[test]" && pytest

Tests are offline: real candles in tests/fixtures/, orders on bytx.testing paper accounts. CI runs them on every push and daily against the newest bytx; before a bytx release, run them against the release candidate: gh workflow run tests.yml -R byttrade/bytx-skills -f bytx=git+https://github.com/byttrade/bytx@<branch>.

Fine print

Not financial advice. Most people who trade actively lose money; leverage makes it faster. A stop is a plan, not a guarantee: gaps, slippage and auto-deleveraging can fill it far away, and liquidation can come first if leverage is high. Funding is charged every period while a position is open. The time machine is a simulation from the last few weeks, not a prediction. Paper trading simplifies margin and liquidation and does not watch your stop between runs. Keys live in environment variables, never in the chat.

MIT. Built on bytx (BYTtrade Source-Available License, installed from PyPI, not copied here). No-code bots, backtests and trade confirmations in Telegram — at byttrade.com.

All risk is yours. bytx-skills is a tool, provided as is, without any warranty. You alone decide whether and how to use it, and you alone bear every loss, missed profit or other consequence, including those caused by errors in the software, the data, an exchange or an AI assistant. Test it yourself — on paper, then on an exchange demo account — and only once you are satisfied that it works for you, use it with real money, in amounts you can afford to lose. The authors and BYTtrade are not liable for any trading results.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI assistants to perform cryptocurrency trading analysis and execution with 38+ tools including real-time market data, technical indicators, risk management, and support for both paper trading and live execution on Hyperliquid.
      7
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI assistants to securely trade on Hyperliquid perpetual exchange, including order placement, position management, market data retrieval, and vault operations via natural language.
      29 PyPI
      24
      MIT
    • A
      license
      B
      quality
      B
      maintenance
      Enables AI assistants to interact with a complete trading infrastructure, including live futures bots, real market data, backtesting, and copy-trading signals. Supports 503 tools across 20 domains, from market data to order execution, with zero synthetic data.
      168
      1
      MIT