Mangrove for AI agents

Mangrove gives trading agents guardrails. Every order goes through built-in risk management: circuit breakers, drawdown and daily loss limits, automatic risk-based position sizing, and automatically generated stop-loss and take-profit orders. The hosted MangroveAI MCP service at https://api.mangrovedeveloper.ai/mcp/ gives an agent more than 100 tools for strategy development, backtesting and live execution across more than a dozen exchange venues. The open-source mangrove-agent is the local client that adds wallets, Kraken, Robinhood and DEX execution on the user’s machine. Connected clients request operations and preserve Mangrove’s decisions, errors, and audit details.

What your agent can do

• Research markets with MangroveAI tools for OHLCV and market data, trending assets, token holders, chain TVL, smart-money sentiment, and social sentiment.
• Develop strategies with Mangrove tools to compose or build candidates, verify and save them, then run backtests and read trades or benchmarks. Follow the current strategy instructions and tool schemas.
• Run paper schedules and request supported live execution actions through Kraken, Robinhood, or DEX workflows when available. Each workflow has its own authorization path and Mangrove-owned risk controls.
• Search prior sweep results and screen candidates before requesting heavier compute. Tool availability and cost depend on the configured service.
• Search the trading knowledge base with kb_search and related document tools.

Capabilities vary by hosted service configuration and local agent version. For hosted MCP, use the standard protocol tools/list operation to discover available tools and schemas. The local agent’s list_tools helper describes its own catalog and may include forwarded tools. A fixed tool count or successful catalog read does not confirm access to every operation.

Connect it

• Install Mangrove Agent: choose an install path for Claude Code, Codex, another MCP client, or a local setup. Use one installation path per machine.
• Claude Code plugin: run claude plugin marketplace add MangroveTechnologies/mangrove-agent, then claude plugin install mangrove-agent@mangrove. Follow the plugin prompt to configure the required API key. See the plugin install guide.
• Local MCP client (Codex, Grok, Cursor, or your own agent): clone the repository and run ./scripts/setup.sh. The local agent serves MCP at http://127.0.0.1:9080/mcp/ by default. See the setup guide and guided tutorial.
• Hosted MCP: connect to the MangroveAI Streamable HTTP endpoint at https://api.mangrovedeveloper.ai/mcp/ and use the client’s protocol-level tools/list operation. See the authentication guide for supported access methods.
• REST: the local agent exposes local tools under /api/v1/agent/. The hosted REST API base is https://api.mangrovedeveloper.ai/api/v1. Start with the API reference and the developer hub.
• Python clients: see the MangroveAI Python SDK and MangroveMarkets Python SDK package pages for their current installation and capability details.

Pay the way an agent pays

Authentication and pricing depend on the service and operation. Some data and backtest calls may be metered, and some workflows support x402 payments. Read the current access documentation and inspect available pricing information before requesting paid or compute-intensive work. Do not assume an operation is free when pricing metadata is absent.

Custody and safety

• Custody depends on the workflow. In Kraken bring-your-own-key mode, the API key stays on the user’s machine and the local client connects directly to Kraken. In Kraken OAuth mode, Mangrove’s platform holds the authorized grant and submits supported operations.
• Local wallet workflows keep key generation, storage, and signing on the user’s machine. A routing service receives a signed transaction, not the private key. Managed organization wallets have a separate custody model.
• Mangrove-owned execution software enforces deterministic risk management based on live market and account conditions. Depending on the workflow, checks may run in a hosted MangroveAI engine or a local Mangrove execution layer. Connected clients must preserve those decisions and errors, not replace or bypass the checks.
• Use the documented authorization and confirmation steps for each workflow. Backtests are historical simulations, not forecasts of returns, and Mangrove does not provide financial advice.

Watch it work

• Mangrove x Robinhood: live trading with Claude Code and Codex (2:26)
• Mangrove MCP playlist: install, ask a market question, build a strategy

For builders

Public developer references: developer hub, OpenAPI specification, quick overview, full guide, and execution configuration authoring schema. The full guide explains the distinction between authoring guidance and accepted transport inputs.

Mangrove Agent is MIT licensed and provides a local client for Mangrove’s documented APIs and MCP services. Review the source repository, developer documentation, API reference, and trading knowledge base before integrating. Current Python package details are on PyPI and PyPI. Questions: Discord.

Listed on