RedoxNet.Mcp.LsOpenApi 1.6.0

RedoxNet.Mcp.LsOpenApi

MCP server for the LS 증권 OpenAPI — exposes Korean and US/overseas (Nasdaq / NYSE / AMEX) stock market data as MCP tools so AI assistants can query quotes, charts, ETF data, market screeners, and index / industry / theme context in natural language, plus the user's live LS broker account (read-only inquiry: holdings, balance, orders, P&L, BEP, orderable capacity) and a local-only paper-portfolio module (multi-broker holdings, watchlists, watched themes, JSON backup / restore).

Unofficial third-party MCP server. Not affiliated with or endorsed by LS Securities Co., Ltd. (LS증권). v1.6 scope: read-only market data + read-only LS broker inquiry + local paper portfolio notes (manual entry; no order placement — that's v1.7).

Install

Prerequisite. dnx is the dotnet tool launcher that ships with .NET SDK 10 or later. Install from .NET downloads if you don't have it yet. Verify with dnx --help.

dnx fetches the latest published version from NuGet on every launch — no separate install step. Wire it into your MCP host:

MCP standard config — Claude Desktop / Claude Code / Cursor / Google Antigravity

These hosts share the same mcpServers JSON schema. Paste the block below into the matching config file and restart the host.

{
  "mcpServers": {
    "lsopenapi": {
      "command": "dnx",
      "args": ["RedoxNet.Mcp.LsOpenApi", "--yes"],
      "env": {
        "LS_APPKEY": "...",
        "LS_APPSECRETKEY": "...",
        "LS_MARKET": "real"   // default if omitted; use "virtual" only for LS mock accounts
      }
    }
  }
}

Among these hosts, the inline chart surface (Plotly v5 spec rendered in chat) lights up on every host that advertises the SEP-1865 io.modelcontextprotocol/ui capability — empirically verified on Claude Desktop / Claude Cowork (and reported by Cursor's changelog). Google Antigravity currently advertises no UI capability, so the server classifies it as text-only: include_chart drops out of the tool schema, the chart spec is stripped, and _meta.render_status: "stripped_text_only" steers the model into honest narration (no false "I drew the chart" claim).

Codex CLI

%USERPROFILE%\.codex\config.toml (Windows) or ~/.codex/config.toml (macOS / Linux):

[mcp_servers.lsopenapi]
command = "dnx"
args = ["RedoxNet.Mcp.LsOpenApi", "--yes"]

[mcp_servers.lsopenapi.env]
LS_APPKEY = "..."
LS_APPSECRETKEY = "..."
LS_MARKET = "real"  # default if omitted; use "virtual" only for LS mock accounts

VS Code

Workspace .vscode/mcp.json — VS Code uses a servers top-level key (not mcpServers) and an explicit "type": "stdio":

{
  "servers": {
    "lsopenapi": {
      "type": "stdio",
      "command": "dnx",
      "args": ["RedoxNet.Mcp.LsOpenApi", "--yes"],
      "env": {
        "LS_APPKEY": "...",
        "LS_APPSECRETKEY": "...",
        "LS_MARKET": "real"  // default if omitted; use "virtual" only for LS mock accounts
      }
    }
  }
}

FieldCure AssistStudio

Settings → Connect → Add MCP Server, then fill the dialog:

LS_APPKEY=...
LS_APPSECRETKEY=...
LS_MARKET=real

AssistStudio renders the optional Plotly chart spec from ls_get_chart / ls_get_overseas_chart inline in the chat — call either with include_chart=true (single timeframe) to get a candlestick chart directly in the conversation. Inline chart rendering also works on Claude Desktop Chat, Claude Cowork, VS Code Chat, and any other SEP-1865 host that advertises the io.modelcontextprotocol/ui capability (v1.5 verified end-to-end). Text-only hosts (Codex, Claude Code CLI) receive _meta.render_status: "stripped_text_only" on the response so the model can narrate honestly that no chart was shown.

Environment variables

Credentials are accepted only through the process environment — never through chat, tool arguments, or MCP elicitation. Prompting for them in conversation would either log them or train callers to share them in transcripts, so that input path is intentionally closed off.

Local data lives at %LOCALAPPDATA%\RedoxNet\LsOpenApi\ on Windows and ~/.local/share/redoxnet/lsopenapi/ on Linux/macOS: token.db (auth cache, SHA-256 keyed) and portfolio.db. The portfolio file carries two logically distinct tables: the paper-portfolio surface (accounts / holdings / watchlist_* / watched_themes) — mode-agnostic, user-entered holdings and watchlists, never auto-synced — and the live LS broker registry (ls_accounts) — auto-populated from the AcntNo / branch / account-name echoed on the first successful ls_account_* call, keyed by (account_no, mode), used purely to label _meta.account_used. Live balances / orders / holdings themselves are never cached; every ls_account_* call is a fresh REST snapshot.

Tools (50 in the standard profile, 53 in all)

Recent surface highlights (full history in RELEASENOTES.Mcp.md):

• v1.6 — LS broker account inquiry family + schema-split live registry. Ten new ls_account_* tools wrap the LS account-inquiry TR family (holdings t0424, orders t0425, balance CSPAQ12200, BEP CSPAQ12300, credit_limit CSPAQ00600, max_order_qty CSPBQ00200, order_history CSPAQ13700, transactions CDPCQ04700, performance FOCCQ33600, daily_pnl t0150/t0151) — read-only inquiry against the appkey-bound LS broker account, no caching, no daemon. Order placement stays explicitly out (v1.7 will ship ls_place_order / ls_cancel_order / ls_amend_order with paper-trading default, preview-required gating, and idempotency tokens). portfolio.db now keeps paper portfolios (accounts — broker is purely a display label) and the auto-discovered LS live row (new ls_accounts table) in physically separate stores so a paper "LS증권" label can never shadow the live broker echo. ls_account_* tools take no account parameter — LS REST routes through the authenticated session, not a client-side selector. New ls_account(action="set_live_nickname") for friendly labelling.
• v1.5 — Fidelity-first chart narration. Every chart-emitting tool response now carries _meta.render_status (delivered / stripped_text_only) so the model has a hard signal whether the host rendered the chart and can narrate honestly. ServerInstructions forbids self-synthesis fallbacks regardless of render_status — no rendering raw OHLCV in Python / JavaScript / PNG, no forwarding the spec to a generic visualize MCP. Chart customization is constrained to ls_add_indicator / ls_reframe_chart; layout-level requests (panel height, sizing) are identified as host panel constraints. output_mode=export responses ship _meta.data_purpose: "analysis_only" + _meta.do_not_render. Tool surface unchanged.
• v1.4 — (a) date envelope (query_date input + data_as_of / query_date_resolution response fields) so non-trading-day fallbacks are explicit — wired on ls_get_market_funds_trend and ls_get_short_selling_trend; (b) LS Q-Click signal screeners: ls_list_screeners, ls_run_screener, ls_combine_screeners over t1825 / t1826.
• v1.3 — first-class US/overseas stocks (Nasdaq / NYSE / AMEX): ls_search_overseas_stock, ls_get_overseas_quote, ls_get_overseas_chart. ls_add_indicator / ls_reframe_chart accept overseas dataset_ids through the shared handle cache.
• v1.2 — chart side-channel adapts per host via MCP Apps capability negotiation (SEP-1865 io.modelcontextprotocol/ui + a clientInfo allowlist). Text-only hosts get neither include_chart nor structuredContent.chart; chart-rendering hosts get the Plotly spec.
• v1.1 — program-trading slice (ls_get_program_trading, ls_analyze_program_flow).
• v0.10 — surface compressed via five action-routed portfolio dispatchers (ls_account, ls_watchlist, ls_watched_themes, ls_portfolio_io, ls_holding); LS_TOOL_PROFILE hides the 3 catalog tools by default.

Market data (LS-backed, credentials required)

ls_search_tr / ls_describe_tr / ls_call_tr are catalog tools — hidden in the default standard profile; set LS_TOOL_PROFILE=all to expose them.

US / overseas individual stocks (LS-backed)

Index + industry (LS-backed)

LS themes (LS-backed)

Screeners & per-stock analytics (LS-backed)

Program trading (LS-backed)

LS broker account inquiry (live, LS-backed)

Fresh REST snapshot of the appkey-bound LS broker account on every call — no caching, no daemon, no background sync. None of these tools accept an account argument: LS account-inquiry TRs route through the authenticated session, not a client-side selector. The first successful call auto-discovers the broker's AcntNo into a separate ls_accounts registry; a friendly nickname can be attached via ls_account(action="set_live_nickname").

Paper portfolio (local-only, no broker sync)

Manual entries persisted to portfolio.db next to token.db — physically separate from the live ls_accounts table above so a paper "LS증권" label never shadows the live broker echo. broker on paper-portfolio rows is purely a display label ("한투", "유안타증권", "LS증권" all valid).

Ambiguity policy (paper). Reads fall back; writes require an explicit target when ambiguous. 0 paper accounts → RequiresAccount; 1 paper account → auto with applied_to echo; 2+ → AmbiguousAccount with candidates[]. Every paper-portfolio mutation includes applied_to (single account) or applied_to[] (corporate actions).

Error envelopes. RequiresAccount / AmbiguousAccount / AccountNotFound / RequiresConfirmation / InsufficientQuantity / ValidationError / LiveAccountNotFound — all carry structured fields (candidates, identifier, holding count + market value, etc.) so the LLM can recover automatically.

Full release notes: https://github.com/redoxnet/mcp-lsopenapi/blob/main/RELEASENOTES.Mcp.md

Documentation & source

• Project home: https://github.com/redoxnet/mcp-lsopenapi
• TR inventory: https://github.com/redoxnet/mcp-lsopenapi/blob/main/docs/LS-TR-INVENTORY.md
• SDK package: https://www.nuget.org/packages/RedoxNet.LsOpenApi.Core/
• License: MIT