v1.6.1

Fixed

• get_chart_screenshot end_time - the scroll now frames the window to actually end at the requested time. In 1.6.0 the capture re-pinned to the latest bar (the screenshot ignored the scroll) and could clamp on deeper history; the EA now anchors the capture to the scrolled window, compensates for the screenshot being wider than the chart, and preloads history before navigating. Requires the rebuilt EA.

v1.6.0

Added

• get_chart_screenshot framing - four optional params to frame the capture: scale (0-5 zoom passthrough), bars (approximate visible candle count, EA self-calibrates to the nearest zoom step; mutually exclusive with scale), end_time (UTC timestamp to scroll the window back to; clamps forward, NO_BARS_AT_TIME when older than history), and price_min/price_max (pin the vertical price band). All default to the live chart framing; requires the rebuilt EA.

v1.5.2

Changed

• get_chart_screenshot annotation guidance - documents that a text annotation anchored near current price at the same level as an hline overlaps the hline's right-margin description, and that the temporary capture chart inherits your default template. Docs only; no behavior change.

v1.5.1

Added

• Chart annotations - get_chart_screenshot now takes an optional
  annotations list to mark up the chart before capture: horizontal lines,
  vertical lines, trendlines, free-placed text and corner labels. Roles
  (resistance, support, note, neutral) pick sensible colors, or set
  color directly. Up to 16 per screenshot.
• Markup is drawn on the temporary chart the capture already uses, so it is
  discarded with that chart and never touches your open charts.

Note

• Requires the updated AgentScreenshot.ex5. Older builds keep working for
  unannotated screenshots but ignore annotations.

v1.5.0

Added

• get_chart_screenshot(symbol, timeframe) - native MT5 chart screenshots (Windows).
  Returns a PNG of the actual MetaTrader 5 chart, including your template, indicators, and drawings, so an LLM can read price action and support/resistance visually instead of parsing OHLC numbers.

v1.4.2

Added

• Prepared for publishing on the official MCP Registry. Updated server.json metadata
  (io.github.vincentwongso/mt5-trading-mcp) and added the mcp-name ownership
  marker to the README so the package can be published to
  registry.modelcontextprotocol.io. No code changes.

v1.4.1

Fixed

• HTTP transport no longer crashes when [transport.http] auth_token is set.
  Startup called mcp.add_middleware(...), but the mcp-package FastMCP has no
  such method (it only exists on the separate standalone fastmcp package), so
  any configured token raised AttributeError: 'FastMCP' object has no attribute 'add_middleware' and the server never started. The bearer-auth path now wraps
  FastMCP's own streamable-HTTP ASGI app in a BearerAuthMiddleware and serves
  it through uvicorn the same way FastMCP does internally, so a static
  shared-secret token works without the OAuth machinery FastMCP's built-in auth
  would otherwise require. The empty-token (unauthenticated) path is unchanged.

v1.4.0

Changed

• Stateless HTTP is now the default ([transport.http] stateless = true).
  The streamable-HTTP transport previously kept one server transport per MCP
  session, which the MCP SDK only frees on a clean session close. A client that
  opened a new session per poll and never closed it leaked a transport each time,
  growing the process unbounded - the memory leak seen on long-running VPS / HTTP
  deployments. Stateless mode builds a fresh transport per request and tears it
  down immediately, so nothing accumulates. Trade-off: server-pushed resource
  subscriptions (quotes://, account://, positions://) need a persistent
  session and are inert in stateless mode (tools still work - clients poll). Set
  stateless = false or serve --no-stateless to restore the old behaviour.
• Default log level is now WARNING ([logging] level). An unattended HTTP
  server no longer floods the console with uvicorn's per-request access log or the
  SDK's "Processing request" / "Created new transport" chatter; consent-posture
  and auth warnings still surface. Set INFO / DEBUG for more, via
  [logging] level, the MT5_MCP_LOG_LEVEL env var, or serve --log-level
  (CLI > env > file).

Added

• serve --stateless / --no-stateless and serve --log-level flags.
• MT5_MCP_LOG_LEVEL environment override for the log level.
• examples/vps/install-mt5-mcp-task.ps1 installs a daily-restart companion
  task (-DailyRestartAt, default 03:30; -NoDailyRestart to skip) that
  stops and restarts the server once a day to reclaim memory on Windows VPS
  deployments.

v1.3.2

Changed

• Eager-connect is now the default. The server connects to MT5 at startup on
  the main thread, before entering the transport loop, instead of lazily on the
  first tool call - so stdio clients (Claude Desktop / Claude Code) and HTTP
  agents no longer hit a multi-minute first call while the MetaTrader5
  C-extension initialises on FastMCP's asyncio event-loop thread. The startup
  connect is non-fatal: if the terminal isn't up yet the server still starts and
  falls back to the lazy path. Disable with serve --no-eager-connect or
  [mt5] eager_connect = false; the serve --eager-connect flag still works.
• .mcp.json is no longer committed. It is per-user Claude Code install
  config; register the server yourself with claude mcp add --transport stdio --scope project mt5-mcp -- python -m mt5_mcp serve (see docs/clients.md).
  Cloning the repo still auto-loads the .claude/skills/.

Packaging

• License metadata now uses the SPDX expression license = "MIT" (PEP 639)
  instead of the deprecated table form plus license classifier.
• The published source distribution is leaner: demo media, CI workflows, and
  editor/agent config are excluded from the sdist (the wheel was already clean).

Added

• CI runs ruff check as a lint gate.

Documentation

• Clarified the Linux host-side bridge as a best-effort / unsupported alternative
  to the recommended all-in-one Docker image; fixed broken setup-doc anchors and
  stale version references; switched repo text to ASCII.

v1.3.1

Changed

• Behavior change: the consent gate is opt-in and off by default again.
  policy.auto_approve_notional reverts to its pre-1.3.0 semantics - the default
  0 now disables the gate, so place_order / close_position /
  modify_order auto-execute (full-open) with no approval step. Set
  auto_approve_notional > 0 to arm the gate: orders/closes at or above that
  notional then require human approval, and widening or removing a stop does too.
  This reverses the 1.3.0 fail-closed default, which broke unattended agents that
  expected orders to execute. The pre-flight limits (max_*) and symbol
  allow/deny lists remain opt-in (0 / empty = off), as before.

Added

• A startup WARNING (both transports) when the consent gate is off
  (auto_approve_notional <= 0), so the fail-open posture is visible in the logs.
  Mirrors the existing empty-auth_token warning.

Fixed

• place_order (priced / pending orders) and modify_order (widening or removing
  a stop) no longer turn a quote outage into an INTERNAL_ERROR when the consent
  gate is armed. If approval is required but no live tick is available to render
  the preview, they now refuse gracefully with SYMBOL_NOT_ENABLED (mirroring
  close_position); modify_order also falls back to the position's last-known
  price for widening detection during the outage.
• Corrected the doctor --smoke-trade hint when the gate is armed: to run the
  smoke order you raise auto_approve_notional above it (or set 0), not lower it.

Upgrade note

If you relied on 1.3.0's fail-closed default to force approval on every trade,
you must now set auto_approve_notional > 0 explicitly to keep that behavior;
otherwise mutating calls execute without an approval step.