Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@fix-mcp diagnose BATS session logon rejection"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
AI agents can write FIX messages. They cannot recover a stuck session, validate a NewOrderSingle, or pause a TWAP that is behind schedule. FIX-MCP gives them the trading-ops brain to do all three through explicit MCP tools, approval gates, and auditable evidence.
FIX-MCP is an open-source professional demo for AI-assisted trading operations. It gives Claude, GPT, Gemini, and any MCP-capable agent a controlled tool surface for diagnosing and resolving realistic FIX, OMS, reference-data, venue, and algo incidents while a human operator stays responsible for approval and final control.
The demo runs against a simulated broker-dealer environment. The product thesis is broader: MCP is the right interface for letting LLMs work with real operational systems without giving them magic, unbounded access to the desk.
22 MCP tools for session repair, order triage, reference-data updates, venue state, algo management, scenario scoring, and trace capture.
14 real-world desk scenarios covering 02:05 ET pre-dawn startup through 16:32 ET after-hours dark-pool failures.
Human-led workflow: investigate, approve workbook, run approved recovery, then stress test only after the baseline path is understood.
Mission Control dashboard with case brief, workbook, operator rail, trace, FIX wire, terminal, manual runbook, and copilot panel.
Production-shaped stack: Python MCP server, REST API, Next.js console, PostgreSQL 16, Redis 7, Docker Compose, and async FIX TCP connector scaffolding.
Related MCP server: a2a-trustgate
System prompt: src/fix_mcp/prompts/trading_ops.py is exposed to MCP clients as fix://prompts/trading-ops. The web console mirrors it in src/store/prompts.ts.
Model configuration: the Mission Control copilot currently calls openai/gpt-5.4 through OpenRouter, either through a server-side OPENROUTER_API_KEY or a user-provided key in the Copilot panel.
Agent framework: this repo does not use LangChain or LangGraph. The control boundary is MCP tools, MCP resources, MCP prompts, FastAPI, and the Next.js operator console.
Hallucination controls: scenario facts come from JSON fixtures and live engine state; actions go through typed MCP tools; production-like steps require human approval; every tool call is visible in Trace.
Demo boundary: the included environment is simulated. Do not connect a public demo to live FIX, OMS, reference-data, or monitoring systems.
See System Prompt, Model, and Guardrails for the demo-facing answer to model, prompt, LangChain/LangGraph, and grounding questions.
You are... | FIX-MCP gives you... |
Broker-dealer ops engineer | A working model for AI-assisted incident triage during market hours. |
OMS / EMS vendor | A reference implementation for adding MCP to trading workflows. |
AI builder | A domain-rich tool surface for agents that need to reason about trading operations. |
VC / fintech evaluator | A concrete artifact showing where AI-in-trading infrastructure is heading. |
The desk loads bats_startup_0200.
Incident: BATS Logon is rejected because the counterparty expects sequence 2450, while the session was reset to 1. Eight overnight GTC orders are blocked, two ETF symbols are missing from extended-hours reference data, and IEX is healthy as fallback.
The operator asks the copilot to investigate. The agent uses MCP tools:
list_scenarios Scenario Loaded: bats_startup_0200
check_fix_sessions BATS down; sequence mismatch detected
query_orders ORDER QUERY — 14 order(s) foundThe agent proposes a recovery workbook:
1. Check BATS session
2. Quantify blocked flow
3. Reconnect BATS
4. Reset BATS sequence if needed
5. Load missing ETF symbols
6. Validate orders releasedThe human approves the workbook. Agent Run executes only the approved path:
fix_session_issue BATS reconnect released stuck orders
fix_session_issue BATS reset_sequence accepted
load_ticker BITO loaded
load_ticker GBTC loaded
validate_orders 14 PASS, 0 FAILThen, and only then, the operator uses Stress Lab to inject a sequence-gap event and prove the system pauses, re-triages, recovers, resumes, and records the trace.
That is the operating pattern: baseline first, pressure test second, evidence always.
git clone https://github.com/henryurlo/fix-mcp.git
cd fix-mcp
docker compose up -dOpen http://localhost:3000.
Login with henry / henry, admin / admin, or click Demo Mode.
Service | URL | Purpose |
Mission Control | Trading-ops dashboard and guided demo workflow | |
REST API | MCP tool dispatch, scenarios, system status | |
MCP stdio |
| Direct MCP protocol entry point |
No Node or Python is required on your host for the Docker demo.
python -m pip install -e ".[dev]"
python -m fix_mcp.apiIn another terminal:
npm install
npm run devThe Next.js console reads BACKEND_URL, defaulting to http://127.0.0.1:8000.
For a local MCP client that can launch the Python entry point:
{
"mcpServers": {
"fix-mcp": {
"command": "fix-mcp"
}
}
}For an HTTP-capable MCP client/proxy:
{
"mcpServers": {
"fix-mcp": {
"command": "npx",
"args": ["-y", "@anthropic-ai/mcp-remote@latest"],
"env": {
"MCP_URL": "http://localhost:8000/mcp"
}
}
}
}Component | Demo | Production / Consulting Engagement |
FIX sessions | Simulated Python objects | Real FIX engine logs and session controls |
OMS | In-memory order state | OMS database/API integration |
Reference data | Preloaded JSON | Vendor feeds, DTCC data, internal symbology |
Monitoring | Scenario engine preloads incidents | Datadog, Splunk, Grafana, or internal event streams |
Execution | Updates simulated state | Sends approved FIX messages or calls approved OMS APIs |
MCP tools | Same tool surface | Same interface, production adapters |
Domain intelligence | Same prompts and logic | Tuned to client workflows, venues, and controls |
The professional work is the integration layer: wire the same MCP interface and trading-ops knowledge into a firm's real logs, OMS, reference data, monitoring, and approval workflow.
MIT. See LICENSE.
Advance the simulated market clock by N minutes. Auto-pauses on triggers.
Inputs: minutes (required).
Submit a structured approval payload for an action. Enforces idempotency and risk-flagged audit trail.
Inputs: action (required), order_ids (required), approved_by (required), risk_flag, +2 more.
Cancel an active algo and send OrderCancelRequest for all open child slices.
Inputs: algo_id (required), reason.
Cancel or replace an existing order via FIX 35=F or 35=G.
Inputs: order_id (required), action (required), new_venue, new_quantity, +2 more.
Check status of algo orders: schedule deviation, execution quality, IS shortfall, over-participation, and child order health.
Inputs: algo_id, symbol, status.
Check FIX session health: status, sequence numbers, heartbeat age, latency.
Inputs: venue.
Report per-symbol market-data staleness in ms. Flags symbols whose last quote exceeds a 500ms advisory threshold. Pass 'symbol' to filter.
Inputs: symbol.
List all orders in pending_ack status with pending age, venue ACK delay, and a duplicate-risk flag (true when pending age > 30 s). Pass 'venue' to filter to a single venue.
Inputs: venue.
Look up a symbol or CUSIP. Returns full record, pending corporate actions, and affected open order count.
Inputs: symbol, cusip.
Clear any simulated feed delay for a venue and report what delay was removed. Returns NOOP when no delay was set.
Inputs: venue (required).
Compare a saved snapshot against current state. Returns structured diff.
Inputs: snapshot_id (required).
Return full session diagnostics for a venue including sequence numbers, latency, heartbeat age, and associated orders.
Inputs: venue (required).
Resolve a FIX session issue: resend_request (gap recovery), reset_sequence, or reconnect.
Inputs: venue (required), action (required).
Get the manual commands equivalent to MCP tools. Filter by tool name.
Inputs: tool.
Get the tool execution trace log with optional filters.
Inputs: limit, tool, status, source.
Get summary statistics for the current trace.
Inputs: no named parameters.
Search log files for a pattern and return matching lines.
Inputs: pattern (required), file (required).
Inject a simulated market event into the current scenario for training. Supported: venue_outage, luld, reject_spike, client_message, seq_gap, sla_breach.
Inputs: event_type (required), target, delay_sec, details.
List all available trading scenarios or load one into the runtime. Use action='list' to see all scenarios with their context summaries. Use action='load' with scenario_name to switch the active scenario.
Inputs: action, scenario_name.
List all saved state snapshots for the current scenario.
Inputs: no named parameters.
Load a new symbol into the reference store and release orders pending that symbol.
Inputs: symbol (required), cusip (required), name (required), listing_exchange (required), +2 more.
Modify an active algo: pause, resume, or update POV participation rate.
Inputs: algo_id (required), action (required), new_pov_rate.
Query OMS orders with optional filters. Returns order details including notional value and SLA countdowns for institutional orders.
Inputs: client_name, symbol, status, venue, +1 more.
Release stuck orders, re-checking blockers. Optional reason_filter limits scope (e.g., 'stale_md', 'venue_down').
Inputs: reason_filter.
Reset FIX sequence numbers for a venue session.
Inputs: venue (required).
Resume a paused simulation after reviewing the trigger.
Inputs: notes.
Restore OMS state from a previously saved snapshot.
Inputs: snapshot_id (required).
Flagship pre-market health check: sessions, corp actions, stuck orders, SLA deadlines, validation summary.
Inputs: no named parameters.
Save a deep-copy snapshot of the current OMS + session state. Use before irreversible actions.
Inputs: label (required).
Compute and return the full KPI score report for the current scenario.
Inputs: no named parameters.
Submit a new algorithmic order (TWAP, VWAP, POV, IS, DARK_AGG, ICEBERG). Creates a parent algo order with execution schedule and initial child slices.
Inputs: symbol (required), side (required), quantity (required), algo_type (required), +6 more.
Send a new order via FIX NewOrderSingle. Validates symbol, checks corp actions, auto-routes if no venue supplied.
Inputs: symbol (required), side (required), quantity (required), order_type (required), +3 more.
Send a heartbeat and return the heartbeat status for a specific venue session.
Inputs: venue (required).
Return the last N lines of a log file.
Inputs: file (required), lines.
Return current simulated time, pause state, and trigger history.
Inputs: no named parameters.
Rename a symbol and bulk-update all open orders. Flags stop orders for manual review.
Inputs: old_symbol (required), new_symbol (required), reason (required).
Change venue status (active/degraded/down). Use fix_session_issue to recover a degraded/down venue.
Inputs: venue (required), status (required).
Validate a set of orders: check symbol validity, venue status, duplicate ClOrdIDs, and client status.
Inputs: order_ids, symbol, status.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 38 tools
Each tool targets a distinct action or resource, from order management to session health and scenario controls. No two tools have overlapping purposes; even similar actions like cancel_algo and cancel_replace are clearly different (algo vs. single order).
All tool names follow a consistent verb_noun pattern in snake_case, e.g., advance_time, approve_action, check_algo_status, diff_snapshot, fix_session_issue. The naming is predictable and easy to understand.
38 tools is significantly above the typical 3-15 range for a well-scoped MCP server. While the domain of trading simulation is complex, the count feels excessive and may overwhelm an agent, reducing coherence.
The tool surface covers the core workflows: order lifecycle (send, cancel, replace, modify), algo orders, session management, scenario simulation, logging, and validation. Minor gaps exist (e.g., no direct market data subscription tool), but most essential operations are present.
Trading MCP server for AI agents, with live market data, account reads and controlled execution.
AI-powered incident management and server monitoring via MCP.
MCP server for Mudrex futures trading enabling AI agents to securely access data and risk tools.
MCP server for your apps' tools and custom tools, plus hosted AI agents and approval-gated workflows
