<div align="center">

AlgoVesta MCP Server

Trade 16 crypto exchanges and MetaTrader 5 from your AI assistant — over one MCP connection.

Documentation · Tool reference · Safety model · Tool schemas (JSON) · Product overview

</div>

---

AlgoVesta runs a hosted Model Context Protocol server that reaches 16 crypto exchanges and MetaTrader 5 through a single HTTPS endpoint. Claude, ChatGPT, Cursor, Claude Code, Gemini CLI or any other MCP-capable client can read balances, open and close positions, move stop-loss and take-profit, and audit its own actions.

Your risk rules are compiled to JSON and evaluated on the server, after the request leaves the model — outside the reach of anything the model can be persuaded to say. Every new key starts on a $5,000 paper balance, and every action returns an ed25519-signed, hash-chained receipt.

Hosted and closed-source. This repository is the public documentation, tool reference and client configuration set. There is nothing to install, build or self-host — you connect to the hosted endpoint.

Contents — At a glance · How it works · Quick start · Connect your client · Supported AI clients · In use · Tools · Safety model · Venues · Performance and limits · Endpoints · FAQ

At a glance

| | | |---|---| | Transport | Remote MCP over Streamable HTTP | | Authentication | Secret link, or OAuth 2.1 with mandatory PKCE (S256) and Dynamic Client Registration (RFC 7591) | | Tools | 20 — 13 read-only, 3 write, 4 destructive (reference) | | Venues | 16 crypto exchanges · MetaTrader 5 · built-in paper engine | | Default scope | paper — $5,000 virtual balance, identical toolset | | Live scope | Issued only after a second factor, enforced server-side | | Risk policy | Compiled to JSON, evaluated deterministically on the server | | Audit | ed25519-signed, hash-chained receipt per action, independently verifiable | | Install | None. Hosted endpoint. | | Registry | com.algovesta/trading — server.json is that manifest |

How it works

flowchart LR
    A["AI client<br/>Claude, ChatGPT, Cursor, and more"]
    B["AlgoVesta MCP server<br/>scope, rate limit, idempotency"]
    C{"Policy wall<br/>evaluated server-side"}
    R["Refusal<br/>plus audit entry"]
    D["Execution"]
    E["16 crypto exchanges"]
    F["MetaTrader 5<br/>managed terminals"]
    G["Paper engine<br/>5,000 USD virtual"]
    H["ed25519-signed receipt"]

    A -->|MCP over HTTPS| B
    B --> C
    C -->|violates a rule| R
    C -->|allowed| D
    D --> E
    D --> F
    D --> G
    D --> H

The model is the caller, never the authority. It decides what to ask for; the server decides what happens. Your exchange API keys never leave AlgoVesta and are created without withdrawal permission — the assistant only ever holds a scoped, revocable MCP link.

Quick start

1. Create an AlgoVesta account and open the MCP Connection tab in the panel.
2. Generate a key. New keys default to the paper scope; the full link is shown once.
3. Paste the link into your AI client as a custom MCP server.

https://api.algovesta.com/u/avmcp_<your-key>/mcp

That URL is a credential. Treat it like a password. If it leaks, revoke it in the panel — revocation takes effect immediately for new connections and drops open event streams within seconds.

Clients that prefer a full authorization flow can use OAuth 2.1 instead, at https://api.algovesta.com/mcp. PKCE (S256) is mandatory and Dynamic Client Registration is supported, so most clients configure themselves. See docs/AUTHENTICATION.md.

Connect your client

Six walkthroughs below; anything else uses the same URL in whatever field it calls a remote MCP server.

Claude — web, desktop, iOS, Android

Settings -> Connectors -> Add custom connector, paste the URL, then allow the tools when Claude asks. Walkthrough

Claude Code

claude mcp add --transport http algovesta https://api.algovesta.com/u/avmcp_<your-key>/mcp

Walkthrough

ChatGPT

Settings -> Connectors -> Advanced -> Developer mode -> Create, transport Streamable HTTP. Custom connectors are a paid-plan feature on OpenAI's side. Walkthrough

Cursor

~/.cursor/mcp.json, or .cursor/mcp.json inside a project — file

{
  "mcpServers": {
    "algovesta": { "url": "https://api.algovesta.com/u/avmcp_<your-key>/mcp" }
  }
}

VS Code — GitHub Copilot

.vscode/mcp.json — file

{
  "servers": {
    "algovesta": {
      "type": "http",
      "url": "https://api.algovesta.com/u/avmcp_<your-key>/mcp"
    }
  }
}

Gemini CLI

~/.gemini/settings.json — file. CLI only; the Gemini web app does not accept custom MCP servers.

{
  "mcpServers": {
    "algovesta": { "httpUrl": "https://api.algovesta.com/u/avmcp_<your-key>/mcp" }
  }
}

Anything else

The three config shapes you will meet, plus a curl connection check: examples/other-clients.md

Supported AI clients

This is a standard remote MCP server over Streamable HTTP, so anything that speaks remote MCP can connect. The clients below document that support themselves — each row links to the client's own configuration docs, which is where the exact field name lives.

<details> <summary><b>Assistants and chat apps</b> — 8</summary>

| Client | MCP documentation | |---|---| | Claude — web, desktop, iOS, Android | Custom connectors | | ChatGPT — Developer mode connectors | OpenAI MCP docs | | Microsoft Copilot Studio agents | Extend with MCP | | Goose (Block) | block.github.io/goose | | LibreChat | MCP in LibreChat | | Open WebUI | docs.openwebui.com | | Cherry Studio | docs.cherry-ai.com | | Raycast | raycast.com |

</details> <details> <summary><b>Coding agents and IDEs</b> — 19</summary>

| Client | MCP documentation | |---|---| | Claude Code | Claude Code MCP | | OpenAI Codex CLI | Codex MCP | | Gemini CLI | MCP servers in Gemini CLI | | Cursor | Cursor MCP | | VS Code — GitHub Copilot | MCP servers in VS Code | | Windsurf | Cascade MCP | | Zed | Zed MCP | | Cline | Connecting to a remote server | | Roo Code | Using MCP in Roo | | Kilo Code | Using MCP in Kilo Code | | Continue | MCP deep dive | | JetBrains AI Assistant | MCP in JetBrains IDEs | | Warp | Warp MCP | | Amazon Q Developer | MCP with Amazon Q | | Kiro (AWS) | Kiro MCP | | Sourcegraph Amp | Amp manual | | Trae | Trae MCP | | PostHog Code | posthog.com/code | | Archestra.AI | archestra.ai |

</details> <details> <summary><b>Frameworks, automation and developer tools</b> — 6</summary>

| Client | MCP documentation | |---|---| | OpenAI Agents SDK | MCP in the Agents SDK | | fast-agent | evalstate/fast-agent | | n8n — MCP Client Tool node | docs.n8n.io | | Postman | MCP requests | | MCP Inspector | modelcontextprotocol/inspector | | MCPJam | mcpjam.com |

</details>

Scope of that claim. We have verified Claude end to end against this server. The others are listed because they document remote MCP support — not because we individually tested each one. Client support also moves fast, and a given client may gate connectors behind a paid plan (ChatGPT does) or behind its CLI only (the Gemini web app does not accept custom MCP servers). If a listed client misbehaves with us, or one is missing, open an issue.

What it looks like in use

You never call a tool by name. You ask in plain language; the assistant picks the tool.

| You say | The assistant reaches for | |---|---| | "What's in my accounts right now, and how has this month gone?" | get_portfolio_context — every exchange, MT5 account and the paper book in one call — then get_trade_history | | "Where is BTC trading on the exchanges I've connected?" | compare_venues — measured price and, where published, bid/ask spread | | "If I opened 0.05 BTC long on Binance with a 2% stop, what would it cost and would my rules allow it?" | simulate_order — a dry run including the policy verdict | | "OK, open it — stop 2% below, target 4% above." | place_order | | "Move the stop on my ETH position up to break-even." | modify_position | | "Close half of the SOL long." | close_position — partial on crypto, full close on MT5 | | "Never risk more than 1% per trade, max 3 positions, stop me after a 5% daily drawdown." | compile_policy — returns readable JSON you approve | | "Show me the receipt for that last order and verify it." | verify_receipt — signature and hash chain, against a public key you can fetch yourself | | "Run my last 90 days of signals again with a 1% risk cap." | simulate_policy, then get_job_status |

And what it will refuse. "Turn on auto-trading for that strategy" — refused; auto_trade is not a writable field in any scope. "Trade on Kraken for me" when Kraken is not connected — VENUE_NOT_CONNECTED, not a guess. "Open it without a stop" — refused, with no saved default to fall back on.

On a paper key every one of those runs against the $5,000 virtual balance with the identical toolset, so the whole wo