← MCP server and skill grades

tradingview-mcp

atilaahmettaner/tradingview-mcp

A
Type
MCP server
Scanned
Nov 12, 2026
Version
ab4681cdbdd1
Files read
140

This MCP server provides AI assistants with 39 read-only tools for stock, crypto, forex, and futures screening, technical analysis, strategy backtesting, and market news. Standard local installations over stdio are low risk, whereas running the optional HTTP or Docker deployment exposes the server without authentication on the network, allowing anyone who can reach the port to drain configured API quotas or paid proxy bandwidth.

What it can reach

Network
Sends HTTPS requests to public endpoints on TradingView, Yahoo Finance, CoinGecko, and Marketaux (optionally via a user-configured HTTP proxy), and opens an HTTP port if started in streamable-http mode.
Secrets
Reads optional MARKETAUX_API_TOKEN and PROXY_* credentials from environment variables or a local .env file.
Files
Reads bundled exchange symbol lists under coinlist/*.txt and an optional local .env file.

Findings

4 low
  1. LowHTTP transport and Docker deployment run without authentication

    src/tradingview_mcp/server.py:1282

    When started in streamable-http mode—which Dockerfile and docker-compose.yml enable by default on 0.0.0.0 (publishing port 8080 on all host interfaces)—FastMCP serves the /mcp endpoint with no password, bearer token, or API key check. Anyone on a reachable network can invoke all 39 tools, exhaust the daily MARKETAUX_API_TOKEN budget, consume paid Webshare rotating proxy bandwidth if PROXY_* variables are set, or trigger upstream rate limits against the host IP.

    Fix: Configure authentication on FastMCP for streamable-http deployments, and bind docker-compose ports to 127.0.0.1 (for example 127.0.0.1:8080:8000) unless placed behind an authenticating reverse proxy.

  2. LowOptional OpenClaw skill instructs agent to run bash commands without confirmation

    openclaw/SKILL.md:27

    The optional OpenClaw integration skill sets always: true and instructs the agent to execute bash commands immediately without asking for clarification, while referencing function names that do not match trading.py CLI subcommands. If a user follows OPENCLAW.md (which includes a public Telegram bot example with allowFrom: ["*"]) on an agent with bash access, untrusted messages or prompt-injected text could trick the agent into running arbitrary shell commands.

    Fix: Connect OpenClaw through structured MCP tool calls rather than a bash wrapper, or restrict the agent to a fixed non-shell command allowlist and warn against enabling bash on public messaging channels.

  3. LowCodex plugin configuration runs an unpinned PyPI package

    .codex-mcp.json:7

    Although .codex-plugin/plugin.json declares version 0.9.0, .codex-mcp.json invokes uvx with tradingview-mcp-server without a version pin. Whenever uvx creates or refreshes its cached environment, it downloads and executes the latest release from PyPI rather than the version reviewed at install time.

    Fix: Pin the package version in .codex-mcp.json to match plugin.json, such as tradingview-mcp-server==0.9.0.

  4. LowDockerfile pipes unpinned remote install script to shell

    Dockerfile:11

    During Docker image builds, the builder stage downloads https://astral.sh/uv/install.sh and pipes it directly into sh without pinning a specific uv version or verifying a SHA-256 checksum. A compromise of the upstream hosting during a build would execute arbitrary code and bake the result into the runtime image.

    Fix: Copy the uv binary from a digest-pinned official container image or download a versioned release archive and verify its SHA-256 checksum before running it.