Best AI Agent Skills for Crypto and Algorithmic Trading 2026

Search a skills directory for crypto or trading and the top results are ccxt, python-binance and AI-Trader. All three are excellent projects. None of them is an agent skill. They are a Python and JavaScript library, another Python library, and a research platform, and installing them teaches your agent nothing.

The list of things that do plug into Claude Code or another MCP client is much shorter. Here it is, with every install command checked against the project's own README rather than a directory listing, on 11 August 2026.

The honest inventory

Stars and dates from the GitHub API on 11 August 2026. Where a licence reads "check the repo", GitHub does not classify one, and I am not going to assert what it is.

The one that is actually a skills repository

tradermonty/claude-trading-skills is the only substantial collection of true agent skills in this space, and it is aimed at equity traders rather than crypto natives. Roughly forty skills organised into market regime, core portfolio, swing opportunities, trade planning, trade memory, strategy research and an advanced satellite group.

The individual skills are more specific than the category names suggest. There is a Follow-Through Day detector built on William O'Neil's methodology, an IBD-style distribution day monitor that tracks 25-session expiry and 5 percent invalidation, a Commitment of Traders contrarian detector implementing step one of Jason Shapiro's approach, and a crypto regime analyser that scores market health 0 to 100 from six components using free keyless public data. Whether you agree with those methodologies is a separate argument. The point is that these encode a named, checkable method rather than "analyse this stock".

Install for Claude Code is a clone and copy, not a package manager command:

• Clone or download the repository
• Copy the skill folder you want, for example backtest-expert, into ~/.claude/skills/ for personal use or .claude/skills/ inside a project
• Restart Claude Code only if you created the top-level skills directory after the session started

For the Claude web app, you download the matching .skill file from skill-packages/, enable code execution and file creation under Settings then Capabilities, and upload the ZIP under Customize then Skills.

The repository is explicit that it is a research, journaling and risk-review assistant and not an automated trading system. It places no orders, runs no hidden scheduled jobs, and keeps human decision gates central. I would trust a trading skill author who says that more than one who does not.

Worth knowing: many of these skills declare required integrations. Several need an FMP key, some need yfinance, some read public CSVs, and a few need chart images. A skill marked as requiring fmp will not do anything useful without that data source.

The MCP servers that connect to something real

Alpaca's server is the only broker-official option here. Alpaca rewrote it for v2 on FastMCP and OpenAPI, and none of the v1 tool names survived, so an older tutorial will not work. Configuration is a straight mcpServers block using uvx alpaca-mcp-server with ALPACA_API_KEY and ALPACA_SECRET_KEY as environment variables. There is no .env file and no init command. Alpaca does not run a hosted remote server, so mobile or ChatGPT use means self-hosting it on a cloud provider and adding it as a custom connector.

TradingView MCP is the most capable market-data option and the one that needs the clearest disclaimer. It is an independent open-source project, not affiliated with or endorsed by TradingView Inc, and it does not log into or scrape a TradingView session. It ships 37 tools covering real-time data, technical indicators, screeners and backtesting across stocks, crypto, forex and futures. Self-hosting is free forever under MIT via pip install tradingview-mcp-server. The author also sells a hosted version at $9 a month for 2,500 calls and 60 a minute, or $29 for 10,000 and 150 a minute, with a three-day trial. Paying for the hosted tier buys you not running Python, which is a reasonable trade for some people and pointless for others.

TradeMemory Protocol does the least glamorous and most useful thing: it gives an agent a decision audit trail and persistent memory across sessions, 20 MCP tools, published to PyPI as tradememory-protocol. If you are letting an agent reason about positions over weeks, the ability to ask what it decided in March and why is worth more than another indicator. Note the repository's default branch is master, which trips up anyone constructing a raw file URL by hand.

Solana MCP by Vybe is a hosted server for live Solana on-chain data and schema browsing. The repository itself is registry metadata rather than the server. The live endpoint is https://mcp.vybenetwork.com, clients authenticate over OAuth, and the config uses npx mcp-remote against the URL with no secrets in the repository. If a directory shows you a .xyz version of that hostname, it is wrong.

Crypto.com's skill is a genuine agent skill for buying, selling, swapping and querying balances through their API, Apache-2.0 licensed. At 16 stars and a last push in June it is early, and it puts trade execution behind a natural-language interface, which brings me to the part most write-ups skip.

The API key problem

Every one of these that can place an order needs a key with trading permission. Handing that key to an autonomous agent is a category of risk that no star count tells you about.

Three rules I would not break.

Use read-only keys unless the agent's specific job is execution. Market data, screening, backtesting and portfolio review all work fine with read-only access. Most of what people actually want from a trading agent needs no trade permission at all.

Start on paper trading. Alpaca supports paper accounts, and there is no reason for an agent's first thousand tool calls to touch real money. If a provider offers no sandbox, that is information about the provider.

Scope and expire the key. Set an IP allowlist where the exchange supports it, disable withdrawal permission always, and rotate on a schedule. A key that can trade but cannot withdraw limits the worst case to bad trades rather than a drained account.

The failure mode here is not a malicious skill, though that is possible and worth auditing for. It is a well-meaning agent misreading an instruction and sizing a position wrong. Deterministic guardrails outside the model, meaning exchange-level limits and permissions, are the only ones that hold when the model is confidently wrong.

The libraries that keep getting mislabelled

Setting the record straight, because these three dominate the search results:

• ccxt, 43,587 stars, MIT. A unified trading API across more than 100 exchanges. Install is npm install ccxt or pip install ccxt. It is a library you write code against, not something an agent loads.
• HKUDS/AI-Trader, 21,233 stars. An agent-native trading research platform from a university lab. Interesting work, not an installable skill.
• sammchardy/python-binance, 7,197 stars, MIT. pip install python-binance. A Binance API wrapper.

All three are worth using if you are writing trading code. None will make your agent better at trading, and a directory that lists them under skills is measuring GitHub stars rather than checking what the repository contains.

What I would install

For research and journaling with no execution risk, start with claude-trading-skills plus TradingView MCP self-hosted. That combination covers screening, technical analysis, regime detection and backtesting for the cost of a pip install, and neither can place an order.

Add TradeMemory Protocol if you are running this over months rather than days.

Add a broker server only when you have a specific execution job, and only on paper first. Alpaca is the one I would pick, because it is published by the broker rather than a third party, and because you can see exactly which tools exist.

For the tooling around this workflow, our full list for finance professionals covers the non-agent side, and our guide to agent skills for data analysts covers the analysis skills that are not trading-specific.

Nothing here is investment advice. These are software recommendations, and the software being good does not make a strategy good.

FAQ

Are there any official broker or exchange agent skills?

Two, and both are early. Alpaca publishes an MCP server for stocks, options and crypto, and Crypto.com publishes an agent skill for trading through its API. Everything else with a broker's name attached is third-party.

Can Claude Code place real trades?

Only if you give it a key with trade permission and connect a server that exposes order tools. Alpaca's and Crypto.com's do. Market-data servers such as TradingView MCP do not, by design.

Is TradingView MCP made by TradingView?

No. The project states plainly that it is an independent open-source project, not affiliated with, endorsed by, or associated with TradingView Inc. It consumes third-party market data and does not automate a TradingView session.

Do I need to pay for TradingView MCP?

No. Self-hosting is free forever under MIT via pip install tradingview-mcp-server. The hosted version at $9 or $29 a month exists so you can skip running Python, and comes with call and rate limits.

Why are ccxt and python-binance listed as skills everywhere?

Because directories rank by GitHub stars and do not check whether a repository contains a SKILL.md. Both are libraries. Check what a row actually is before you install it, no matter how many stars it has.