The trading agent wiki.
An LLM wiki of trading agents. Every skill, MCP server, execution API, and framework for building autonomous trading agents is an interlinked entity page, written for language models and people. Follow the wiki links, or read any page as plain markdown.
Reading this as an agent
- Every entry is an entity page; cross-references are wiki links.
- Each page is raw markdown at
/wiki/<slug>.md. Read it directly, no vector store. - Machine index of every page: /llms.txt. Crawlable graph: /graph.json.
- Concept pages link out to fuller write-ups and primary sources.
Execution APIs
- Superior Trade Managed execution, backtesting, and a deployment lifecycle behind agent-callable endpoints, with an atomic one-call exit and HIP-3 coverage. Closed core, public OpenAPI and docs.
MCP servers
- Alpaca MCP Official MCP server for the Alpaca brokerage API. Strong on regulated equities, lighter on crypto-native primitives.
- edkdev hyperliquid-mcp Community MCP server wrapping the Hyperliquid SDK, with agent-wallet delegation. Solo-maintained, no lifecycle tooling.
- Kraken CLI Single-binary execution engine, 134 commands, built-in MCP server and paper-trading mode. CEX only.
- OKX Agent Trade Kit Official open-source MCP toolkit, ~164 tools across spot, perps, options, and grid bots. Keys stay off the model.
- Senpi Open-source skill library (~65 MCP tools) with strategy archetypes, dynamic stops, and capital isolation per skill.
Agent skills
- Bankr / OpenClaw skills Wallet-centric skills for swaps, limits, DCA, and leverage across Base, Ethereum, Solana and more.
- Dwellir hyperliquid-skills Open skill set for placing and managing Hyperliquid positions from an agent. Thin on backtest/deploy.
- Senpi Open-source skill library (~65 MCP tools) with strategy archetypes, dynamic stops, and capital isolation per skill.
- Superior Trade Managed execution, backtesting, and a deployment lifecycle behind agent-callable endpoints, with an atomic one-call exit and HIP-3 coverage. Closed core, public OpenAPI and docs.
Frameworks
- Freqtrade The most-used open-source crypto bot: build, backtest, hyperopt, dry-run, and run live, with an ML module (FreqAI). Now supports Hyperliquid. Not agent-callable out of the box.
- HKUDS AI-Trader Agent-native trading framework where agents join via messages; strong research/backtest story, lighter execution safety.
- Hummingbot Open-source framework specialized in market making and CEX/DEX arbitrage. Strong live execution; lighter on backtesting and not agent-native.
- NautilusTrader High-performance event-driven engine (Rust core, Python API). Backtest and go live on the same code. Production-grade, but you wire the agent layer yourself.
- TradingAgents Popular open-source multi-agent LLM framework that simulates a trading firm (analysts, researchers, trader, risk debate). Research and backtest focused, not a live-execution layer.
Platforms
- Composer No-code rules-based strategy platform with solid backtest and deploy, but built for humans more than callable agents.
CLI
- Kraken CLI Single-binary execution engine, 134 commands, built-in MCP server and paper-trading mode. CEX only.
Research
- HKUDS AI-Trader Agent-native trading framework where agents join via messages; strong research/backtest story, lighter execution safety.
- TradingAgents Popular open-source multi-agent LLM framework that simulates a trading firm (analysts, researchers, trader, risk debate). Research and backtest focused, not a live-execution layer.
Venues
- Hyperliquid An on-chain perpetuals venue with a fully on-chain order book, sub-second blocks, and gas-free order actions, unusually friendly to autonomous agents.
Concepts
- Agent wallet A delegated key that can place and cancel orders for an account but cannot withdraw funds. If it leaks, the worst case is unwanted trades, not a drained account.
- Backtesting Testing a strategy on historical data. Done honestly (with slippage, fees, and funding) it earns the right to paper-trade, not the right to go live.
- Hyperliquid An on-chain perpetuals venue with a fully on-chain order book, sub-second blocks, and gas-free order actions, unusually friendly to autonomous agents.
- LLM wiki A knowledge base written for a language model: plain markdown entity pages linked with [[wiki-links]] that an agent builds and maintains, so knowledge compounds. Lighter than RAG.
- Loop engineering Designing the control loop an agent runs in (gather context, decide, act, observe, repeat) rather than hand-writing each prompt. A trading agent is a high-stakes example.
- MCP (Model Context Protocol) A standard way to hand an AI model a set of typed tools it can call. For trading, the server holds the keys and signs orders; the model only sees tools.
Ranking
The tools above, scored on one public rubric. Each score (0-5) is our reading of public docs and code as of June 2026, weighted per the methodology. Estimates, not paid placements.
Execution (25%): Order placement, partial fills, and a reliable one-call exit.
Safety (20%): Non-custodial keys, reduce-only, size and loss limits.
Lifecycle (20%): Backtest, paper, deploy, and monitor in one place.
Agent-native (20%): Callable as MCP tools or a typed skill, built for agents.
Openness (15%): Open source, public docs, and an inspectable surface.
| # | Project | Score |
|---|---|---|
| 01 | OKX Agent Trade Kit Official open-source MCP toolkit, ~164 tools across spot, perps, options, and grid bots. Keys stay off the model. | 4.60 / 5 |
| 02 | Superior Trade Managed execution, backtesting, and a deployment lifecycle behind agent-callable endpoints, with an atomic one-call exit and HIP-3 coverage. Closed core, public OpenAPI and docs. | 4.50 / 5 |
| 03 | Senpi Open-source skill library (~65 MCP tools) with strategy archetypes, dynamic stops, and capital isolation per skill. | 4.35 / 5 |
| 04 | NautilusTrader High-performance event-driven engine (Rust core, Python API). Backtest and go live on the same code. Production-grade, but you wire the agent layer yourself. | 4.20 / 5 |
| 05 | Kraken CLI Single-binary execution engine, 134 commands, built-in MCP server and paper-trading mode. CEX only. | 3.95 / 5 |
| 06 | Freqtrade The most-used open-source crypto bot: build, backtest, hyperopt, dry-run, and run live, with an ML module (FreqAI). Now supports Hyperliquid. Not agent-callable out of the box. | 3.75 / 5 |
| 07 | Alpaca MCP Official MCP server for the Alpaca brokerage API. Strong on regulated equities, lighter on crypto-native primitives. | 3.60 / 5 |
| 08 | Hummingbot Open-source framework specialized in market making and CEX/DEX arbitrage. Strong live execution; lighter on backtesting and not agent-native. | 3.35 / 5 |
| 09 | Dwellir hyperliquid-skills Open skill set for placing and managing Hyperliquid positions from an agent. Thin on backtest/deploy. | 3.30 / 5 |
| 10 | HKUDS AI-Trader Agent-native trading framework where agents join via messages; strong research/backtest story, lighter execution safety. | 3.30 / 5 |
| 11 | TradingAgents Popular open-source multi-agent LLM framework that simulates a trading firm (analysts, researchers, trader, risk debate). Research and backtest focused, not a live-execution layer. | 3.25 / 5 |
| 12 | Bankr / OpenClaw skills Wallet-centric skills for swaps, limits, DCA, and leverage across Base, Ethereum, Solana and more. | 3.15 / 5 |
| 13 | edkdev hyperliquid-mcp Community MCP server wrapping the Hyperliquid SDK, with agent-wallet delegation. Solo-maintained, no lifecycle tooling. | 3.10 / 5 |
| 14 | Composer No-code rules-based strategy platform with solid backtest and deploy, but built for humans more than callable agents. | 3.05 / 5 |
FAQ
What is this?
An LLM wiki of trading agents: every skill, MCP server, execution API, and framework as an interlinked entity page, written so a language model can read it directly. Each page is also available as raw markdown, and cross-references are wiki links, so the whole thing reads as a connected knowledge graph rather than a list.
How is the ranking calculated?
A weighted score across five criteria: execution, safety, lifecycle, agent-native, and openness. Weights and definitions are on the methodology page. Each score (0-5) is our reading of the project's public docs and code, not a paid placement.
How does an agent read this wiki?
Fetch any entity page as markdown at /wiki/<slug>.md, or start from /llms.txt for the full index. Because the pages are plain markdown linked with wiki links, an agent can follow references directly without a vector store, which is the point of the LLM-wiki format.
How do I get a tool added or a score fixed?
Open an issue or email us with the repo or docs link and the correction. We re-score against the published criteria and update the page.