trading-brain

# trading-brain Personal trading copilot built as a set of MCP servers for Claude. Connects Claude to **Sierra Chart** (live market data and account state) and **Discord** (a trading-education community used as a knowledge source), with a local SQLite journal for persistent memory. ## What's inside | MCP | Purpose | Tools | |---|---|---| | **sierra-mcp** | Sierra Chart bridge (futures: ES, NQ, MES, MNQ on CME) | `ping_sierra`, `get_quote`, `get_recent_bars` (DTC), `get_recent_bars_scid` (local file), `get_latest_tick_scid`, `get_scid_status`, `get_futures_context`, `get_market_features`, `list_trade_accounts`, `get_account_balance`, `get_positions`, `get_open_orders`, `place_sim_market_order`, `close_sim_position` | | **discord-mcp** | Read-only Discord channels as a knowledge base | `list_servers`, `list_channels`, `read_recent_messages`, `search_messages`, `fetch_image`, `list_groups`, `read_group` | | **journal-mcp** | Local SQLite memory for observations, trades, reviews, and structured playbook knowledge | Tools: `log_observation`, `log_trade`, `update_trade`, `search_journal`, `list_recent`, `daily_summary`, `add_knowledge_item`, `search_knowledge`, `list_knowledge_topics`, `review_knowledge_item`, `promote_observation_to_knowledge`, `list_workflows`, `premarket_review`, `postmarket_review`, `weekly_trading_review`, `discord_study_ingest` | Discord is read-only. journal-mcp writes only to a local SQLite journal database. sierra-mcp includes tightly-guarded simulation/evaluator-only order tools. Real money order placement is out of scope. ## How they talk to Sierra Chart Two paths, complementary: 1. **DTC Protocol** (`get_quote`, `get_recent_bars`, account / positions) — TCP JSON over Sierra's local DTC server. Requires Denali Exchange Data Feed + Service Package 11+ for full market-data redistribution. SC Data (free/delayed) refuses DTC redistribution. 2. **`.scid` file reader** (`get_recent_bars_scid`, `get_latest_tick_scid`) — reads Sierra's local tick-storage files directly from disk, aggregates ticks into bars on the fly, and can read the latest tick with file freshness. Works regardless of DTC permissions. ~Seconds of lag from real-time (file flush cadence). `get_futures_context` is the compact ES/NQ context pack: latest tick, freshness, recent bars, approximate current Globex session OHLC/VWAP/delta, ATR, and a simple bias read. `get_market_features` is the indicator-engine path: it calculates current/previous Globex VWAP, POC, VAH/VAL, high-volume nodes, delta and price location from raw `.scid` ticks. ## Requirements - Windows (Sierra Chart is Windows-native; paths assume `D:\SierraChart\Data\`) - Python 3.11+ - [uv](https://docs.astral.sh/uv/) for env/lockfile management - Sierra Chart account with DTC enabled (Global Settings → Sierra Chart Server Settings → Encoding: **JSON**) - Discord bot token (Application from [discord.com/developers/applications](https://discord.com/developers/applications), Message Content Intent enabled) - Claude Desktop ## Setup ```powershell # Clone git clone https://github.com/AndreuCrespo/trading-brain.git D:\inversion cd D:\inversion # Install each MCP (creates per-project .venv + uv.lock) cd sierra-mcp ; uv sync ; cd .. cd discord-mcp ; uv sync ; cd .. cd journal-mcp ; uv sync ; cd .. ``` Per project, copy `.env.example` → `.env` and fill in secrets. discord-mcp's `.env`: ``` DISCORD_BOT_TOKEN=your-bot-token ``` sierra-mcp env vars (all optional, sensible defaults): ``` SIERRA_DTC_HOST=127.0.0.1 SIERRA_DTC_PORT=11099 SIERRA_DTC_HISTORICAL_PORT=11098 SIERRA_DATA_PATH=D:\SierraChart\Data ``` For order tools, create a local gitignored `sierra-mcp/safety.json` from `sierra-mcp/safety.example.json` and list the demo/evaluator accounts Claude is allowed to touch: ```json { "allowed_sim_accounts": ["SimTB1"] } ``` journal-mcp env vars (optional): ``` JOURNAL_DB_PATH=D:\inversion\journal-mcp\data\journal.db ``` ## Wiring into Claude Desktop Edit `%APPDATA%\Claude\claude_desktop_config.json` and add: ```json { "mcpServers": { "sierra": { "command": "D:\\inversion\\sierra-mcp\\.venv\\Scripts\\python.exe", "args": ["D:\\inversion\\sierra-mcp\\src\\sierra_mcp\\server.py"] }, "discord": { "command": "D:\\inversion\\discord-mcp\\.venv\\Scripts\\python.exe", "args": ["D:\\inversion\\discord-mcp\\src\\discord_mcp\\server.py"] }, "journal": { "command": "D:\\inversion\\journal-mcp\\.venv\\Scripts\\python.exe", "args": ["D:\\inversion\\journal-mcp\\src\\journal_mcp\\server.py"] } } } ``` Restart Claude Desktop fully (Quit from system tray, not minimise). The MCPs appear under Configuración → Desarrollador. > Note: args use the absolute file path instead of `python -m sierra_mcp.server` > because Claude Desktop mangles the `ó` in the project path when invoking > `-m` module mode. Loading the file directly works around this; each > `server.py` has a `sys.path` shim at the top so its package imports resolve. ## Dev workflow (MCP Inspector) ```powershell cd sierra-mcp # or discord-mcp / journal-mcp uv run mcp dev src/sierra_mcp/server.py ``` In the inspector UI, point Command at the per-project venv python and Args at the file path (same reason as Claude Desktop): - Command: `D:\inversion\sierra-mcp\.venv\Scripts\python.exe` - Arguments: `D:\inversion\sierra-mcp\src\sierra_mcp\server.py` ## Discord channel taxonomy discord-mcp groups channels into domain categories defined in [`discord-mcp/groups.yml`](discord-mcp/groups.yml): `prep`, `post`, `recursos`, `fondeo`, `futuros`, `stocks`. Matched by exact name or prefix. Edit the YAML to add new channels — no code changes needed. `read_group` fans out reads across every channel in a group in parallel: ``` read_group(server="...", group="post", limit_per_channel=5) # → snapshot of postmarket-futuros, postmarket-stocks, setups-database, # writes-up all at once ``` ## Journal memory journal-mcp stores local memory in SQLite (default: `journal-mcp/data/journal.db`, gitignored). It is intentionally simple: observations and trades, searchable by text, symbol, tags, dates, and kind. Use it for market observations, trade plans, post-mortems, recurring mistakes, rules, and daily/weekly review raw material. It also has a structured `knowledge_items` table for playbook knowledge derived from Discord/PPT/video/manual review. Knowledge items have lifecycle fields: `status` (`draft`, `reviewed`, `approved`, `active`, `stale`, `deprecated`), `confidence` (`explicit`, `inferred`, `uncertain`), `kind`, `topic`, source references, validity dates and optional `superseded_by`. Trading decision workflows should use `search_knowledge` defaults, which return only `approved`/`active` items. It also exposes workflow tools (and matching MCP prompts) that orchestrate the other MCPs from Claude Desktop: - `premarket_review`: reads Discord `prep`, Sierra market features, and journal playbook/risk memory. - `postmarket_review`: reads Discord `post`, chart screenshots, Sierra market features, and journal memory. - `weekly_trading_review`: summarizes the week from journal + Discord write-ups + current market features. - `discord_study_ingest`: reads Discord course channels/screenshots and saves structured draft playbook knowledge for later review/approval. Example flow: ``` log_observation(content="MES rejected VWAP after postmarket context", tags=["vwap", "post"], symbol="MES") log_trade(symbol="MESM26-CME", side="long", entry_price=7590.5, quantity=1, account="Sim1", setup="EF") daily_summary(date="2026-06-01") ``` ## Example conversations > *"Read the last 5 messages from the `post` group on Estudio trading donAdri, > grab the last 10 1-min bars of MESM26-CME, and tell me whether anything > discussed lines up with recent price action."* > *"There's a 29-May post in postmarket-futuros with two attached screenshots. > Read them and summarise what the EOM trade looked like."* > *"Walk me through `📘15-theplan` and tell me how donAdri's system handles > position sizing."* > *"Log today's main mistake as a journal observation tagged `risk` and > `discipline`, then show me the last 10 journal entries for MES."* > Run the `premarket_review` prompt before the session, then save the final plan > to the journal if it looks right. ## Status & known limitations - ✅ `ping_sierra`, `get_recent_bars_scid` validated against live data - ✅ `get_latest_tick_scid` smoke-tested on MESM26-CME with ~1-2s file/tick age - ✅ `get_futures_context` smoke-tested on MESM26-CME using current Globex session - ✅ `get_market_features` calculates current/previous Globex VWAP, POC, VAH/VAL and value-location tags from local `.scid` ticks - ✅ Discord tools all validated against the live community server - ✅ journal-mcp smoke-tested locally against SQLite - ⚠️ `get_quote` / `get_recent_bars` (DTC) return *"Request is not authorized"* on the current account despite SP11 + Denali active. Workaround in place via `get_recent_bars_scid`; investigating with Sierra support. - ✅ `get_positions` responds correctly after the DTC message ID fix; validate it with a manually-opened Sim position before adding order placement. - ✅ `place_sim_market_order` preview/guards validated. Live submission requires Sierra DTC listening and Trade Simulation Mode On. - ⚠️ `get_account_balance` returns empty on Sim accounts. Sierra may not maintain balances for Trade Simulation Mode accounts. - No real-money order placement. ## Roadmap Short term, in priority order: 1. Use journal-mcp in real Claude Desktop conversations and refine the schema from actual workflow pain. 2. Extend market features on top of `.scid` ticks: RTH/ETH splits, pVWAP/mVWAP, ADR/RVOL and richer market-profile/state tags. 3. If true bid/ask/DOM real-time is required, build an ACSIL bridge inside Sierra Chart that writes ticks/quotes/depth to a local file or socket. 4. Order placement (sim first, with explicit confirmation per call). Longer term: more data sources as separate MCPs (SEC EDGAR, FRED, news), simple backtest framework over `.scid`. ## Layout ``` inversion/ ├── README.md ├── .gitignore ├── sierra-mcp/ │ ├── CLAUDE.md # per-project guidance for future Claude sessions │ ├── pyproject.toml │ ├── uv.lock │ └── src/sierra_mcp/ │ ├── server.py # FastMCP tools │ ├── dtc_client.py # async DTC over TCP/JSON │ ├── dtc_messages.py │ ├── scid_reader.py # local .scid tick file reader │ └── config.py └── discord-mcp/ ├── pyproject.toml ├── uv.lock ├── groups.yml # channel taxonomy (edit me, not code) └── src/discord_mcp/ ├── server.py ├── client.py # Discord REST API client ├── groups.py # YAML loader + channel matcher └── config.py └── journal-mcp/ ├── pyproject.toml ├── uv.lock ├── data/ # local SQLite DBs, gitignored └── src/journal_mcp/ ├── server.py # FastMCP tools ├── db.py # SQLite schema/helpers └── config.py ```