project-lithos
An MCP server for institutional quantitative intelligence, integrating Calera FinSec SEC auditing, ICX infinite context memory, level-2 order book analysis, SQL backtesting, and post-trade execution auditing with security safeguards.
README
🪨 PROJECT LITHOS
Institutional Quantitative Intelligence, Calera FinSec SEC Auditing & ICX Infinite Context Architecture for General LLMs via Model Context Protocol (MCP)
"Unshakeable Ground Truth for High-Velocity Markets: Project Lithos is the solid, deterministic bedrock beneath modern quantitative AI."
🌋 Project Lore & The Vibe Coding Origin
Project Lithos was created during the Live Stream Vibe Coding Stream as an open educational reference harness and pilot testbed for the Calera Labs ICX Infinite Context API and the Calera Labs FinSec (FinanceSec) MCP Server.
In geology, the Lithosphere is the unshakeable outer crust and bedrock of the Earth.
In financial quantitative engineering, combining commercial reasoning LLMs (Google Gemini 3.1 Flash-Lite, Claude 3.5 Sonnet, or GPT-4o) with high-velocity market data and trading infrastructure requires deterministic ground truth and zero hallucination:
- Deterministic SEC Filing Provenance (Calera FinSec MCP — Enterprise-speed $<2\text{ms}$ GAAP extraction, covenants, segment mix, 0% calculation error).
- Infinite Context Memory & Real-Time Delta Sync (Calera ICX A4 Lattice — Sub-millisecond associative recall $<1\text{ms}$, zero token inflation).
- Level-2 Order Book Depth & Squeeze Engine (Real-time bid/ask ladders, order imbalance ratios, liquidity squeeze detection).
- HFT Post-Trade Telemetry Auditing (Pinpointing route bottlenecks, TCP retransmits, and $>50\text{ms}$ latency slippage).
- Topological Contradiction Detection ($Betti-2$ shear voids between supplier warnings and corporate guidance).
- Local 5-Year SQL Alpha Backtesting (Dip-and-rebound patterns, volume accumulation anomalies).
- Cryptographically Guarded Broker Execution (Human-in-the-loop HMAC tokens + Prompt Injection Shield).
🏛️ System Architecture
┌────────────────────────────────────────────────────────┐
│ Commercial Cloud Reasoning LLMs │
│ (Claude 3.5 Sonnet / GPT-4o / Gemini BYOK) │
└───────────────────────────┬────────────────────────────┘
│
▼ ▲ (Secure JSON-RPC 2.0 / stdio / SSE)
│
┌───────────────────────────┴────────────────────────────┐
│ Local MCP Host │
│ (Claude Desktop / Cursor / Antigravity CLI) │
└───────────────────────────┬────────────────────────────┘
│
▼ ▲
┌─────────────────────────────────────────┴─────────────────────────────────────────┐
│ PROJECT LITHOS MCP SERVER │
│ ┌─────────────────────────────────────────────────────────────────────────────┐ │
│ │ 🛡️ SECURITY GUARD & HITL GATEWAY (HMAC Token + Prompt Injection Shield) │ │
│ └──────────┬──────────────┬──────────────┬──────────────┬──────────────┬──────┘ │
└─────────────┼──────────────┼──────────────┼──────────────┼──────────────┼─────────┘
│ │ │ │ │
┌──────────┴─────┐ ┌──────┴────────┐ ┌───┴──────────┐ ┌─┴────────────┐ ┌┴──────────┐
│ Calera FinSec │ │ Calera ICX │ │ Live L2 Book │ │ Local SQL │ │ Post-Trade│
│ Hosted MCP │ │Infinite Memory│ │Depth & Spread│ │ Database │ │ Logs & ACK│
│(SEC 10-K/10-Q) │ │ (A4 Lattice) │ │ (Microstr.) │ │(5-Yr Intrad.)│ │ (>50ms) │
└──────────┬─────┘ └──────┬────────┘ └───┬──────────┘ └─┬────────────┘ └┬──────────┘
│ │ │ │ │
▼ ▼ ▼ ▼ ▼
finsec.caleralabs icx.api.caleralabs Polygon.io / SQLite / Execution
.com/mcp .com/v1 Alpaca Feeds TimescaleDB Logs & CSVs
flowchart TD
subgraph ClientLayer["AI Client & MCP Host Layer"]
LLM["Commercial AI (Gemini BYOK / Claude / GPT)"]
Host["MCP Host (Claude Desktop / Cursor / Antigravity CLI)"]
LLM <--> Host
end
subgraph LithosCore["Project LITHOS Core Engine"]
Server["Project LITHOS MCP Server (server.py)"]
Guard["🛡️ Security Guard & HITL Gateway\n(HMAC Tokens + Prompt Injection Shield)"]
Router["⚡ Autonomous Query Router\n& Slash Dispatcher (chat.py)"]
Host <--> Server
Server <--> Guard
Server <--> Router
end
subgraph Substrates["Integrated Substrates & Data Feeds"]
FinSec["🏛️ Calera FinSec MCP\n(Deterministic SEC 10-K/10-Q)"]
ICX["🌌 Calera ICX\n(4D A4 Infinite Lattice Memory)"]
L2["📊 Level-2 Order Book\n(Imbalance & Squeeze Engine)"]
SQL["🗄️ 5-Year SQL Database\n(Alpha Backtester)"]
Logs["📋 Execution Log Auditor\n(Latency Anomaly Detection)"]
Broker["💼 Alpaca Paper Broker\n(Safe Order Staging)"]
end
Guard --> FinSec
Guard --> ICX
Guard --> L2
Guard --> SQL
Guard --> Logs
Guard --> Broker
⚡ 3-Minute Quickstart
1. Clone & Set Up Environment
git clone https://github.com/Calera-Labs/PROJECT_LITHOS.git
cd PROJECT_LITHOS
# Create and activate virtual environment
python -m venv .venv
# On Linux/macOS:
source .venv/bin/activate
# On Windows (PowerShell):
.\.venv\Scripts\Activate.ps1
# Install dependencies
pip install -r requirements.txt
2. Configure Your API Keys (.env)
Copy the template configuration:
cp .env.example .env
Edit .env with your personal Bring-Your-Own-Key credentials:
# Google Gemini BYOK (Free API key from https://aistudio.google.com/app/apikey)
GEMINI_API_KEY=YOUR_GEMINI_API_KEY_HERE
GEMINI_MODEL=gemini-3.1-flash-lite
BYOK_PROVIDER=gemini
# Calera Labs ICX Gateway
ICX_BASE_URL=https://icx.api.caleralabs.com/v1
ICX_API_KEY=YOUR_CALERA_ICX_API_KEY_HERE
ICX_SPACE_ID=lithos_market_memory
# Calera Labs FinSec MCP (Set to true to run in local deterministic mode)
FINSEC_MCP_URL=https://finsec.caleralabs.com/mcp
FINSEC_API_KEY=STANDBY_FOR_KEY
FINSEC_STANDBY_MODE=true
# Safe Execution Sandbox
LITHOS_BROKER_ENV=paper
3. Run Diagnostic Verification
Run the built-in system verification tool to check your environment, seed the database, and test connectivity:
python verify_setup.py
4. Launch the Interactive Quantitative Terminal
python chat.py
💡 Dynamic Prompt Coordination
The terminal CLI dynamically inspects the health of all three substrates and updates the prompt in real time:
| Prompt Display | Meaning / Active Substrates |
|---|---|
LITHOS [ICX+Gemini] ❯ |
All 3 substrates healthy: Calera FinSec active (indicated by brackets [ ]), Calera ICX active, Gemini BYOK reasoning active. |
LITHOS [ICX] ❯ |
Gemini down/unconfigured: Queries seamlessly execute via Zero-LLM Direct ICX Memory Recall on the A4 lattice. |
LITHOS ICX+Gemini ❯ |
FinSec offline: ICX and Gemini active, brackets omitted. |
LITHOS [Gemini] ❯ |
ICX offline: FinSec and Gemini active. |
LITHOS OFFLINE ❯ |
All offline: Running in local fallback mode. |
🛠️ Complete MCP Tool Catalog (18 Tools)
When running server.py as an MCP server, Project LITHOS exposes 18 tools to your AI agent:
1. Calera Labs FinSec (FinanceSec) SEC Filing Auditing (7 Official Tools)
query_financial_sec(query, low_tokens, usd_only): Certified natural-language SEC EDGAR fact recall with accession provenance or safe refusal.query_sec_metric_exact(company, metric, period): Exact EDGAR metric recall for company + metric + period (gross_margin,total_revenue).query_sec_sector_peers(sector_or_industry, period, limit): Certified XBRL metrics across industry peer groups derived dynamically from SIC codes.compute_sec_cagr(company, metric, start_period, end_period): Compound annual growth rate calculation over verified SEC filing facts with 0% calculation error.lattice_arith_evaluate(op, company, metric, start_period, end_period): Deterministic zero-error algebraic solver (cagr,ratio,multiply,divide,yoy_series).valuation_inputs(company, pack, period): Certified SEC valuation inputs pack (12 packs available:segment_mix,equity_screen,quality_of_earnings,returns_screen,leverage_screen,piotroski_f,altman_z_prime,beneish_m_score,dupont_5step,working_capital_efficiency,dcf_valuation_inputs,ev_bridge).vln_capabilities_overview(): Complete catalog of certified valuation models and arithmetic solvers.
2. Calera Labs ICX Infinite Context Memory Substrate
query_icx_infinite_memory(query, session_id): Grounded reasoning in the Calera ICX Volumetric Lattice Network (VLN) powered by Google Gemini 3.1 Flash-Lite BYOK.sync_market_to_icx_lattice(news_title, text, family): Streams breaking news, transcripts, or filings into the A4 lattice without LLM context token inflation.query_icx_symbolic_ledger(query): Sub-millisecond associative recall ($<1\text{ms}$) for exact numerical invariants.quote_icx_memory_slot(family, index): Deterministic Scoped Recall to quote exact document registers with zero LLM inference.detect_market_contradictions(): Detects $Betti-2$ topological voids and guidance divergence between suppliers and competitors.get_icx_lattice_stats(): Inspects active simplicial nodes, memory columns, crystallized bytes, and 4D A4 manifold geometry.reset_icx_session(session_id): Resets conversational turns while preserving crystallized facts in the underlying lattice space.
3. Live Market Data & Level-2 Order Book Microstructure
get_live_quote(ticker): Real-time prices, day ranges, volume, and VWAP.get_order_book_depth(ticker, depth): Computes bid/ask ladders, order imbalance ratios, and liquidity squeeze conditions.
4. Local Database Auditing & Alpha Backtesting
get_financial_db_schema(): Table schemas and row metrics across local time-series tables.query_financial_db(sql_query): Safe, sandboxed read-only SQL execution (destructive statements blocked).backtest_alpha_pattern(ticker, min_morning_drop_pct, require_green_close): Audits 5-year daily/intraday data for morning dip reversals with volume analysis.
5. Post-Trade Analysis & Execution Telemetry
audit_execution_logs(min_latency_ms, ticker_filter): Isolates orders suffering $>50\text{ms}$ latency/slippage and identifies exchange routing bottlenecks.
6. Systematic Rebalancing & Safe Execution Gateways
get_portfolio_summary(): Holdings, equity, cash, and sector exposure weights.propose_portfolio_rebalance(macro_rationale, tech_shift_pct, fixed_income_shift_pct): Systematic rule-based rebalancer that stages orders with HMAC tokens.execute_broker_order(order_id, confirmation_token): Dispatches trades only after verifying the human confirmation token.
🕹️ Interactive Terminal Commands
Inside python chat.py, use slash commands for instant deterministic operations:
| Command | Syntax Example | Description |
|---|---|---|
| /sec | /sec NVDA gross margin FY2026 |
Query certified SEC EDGAR facts with paragraph provenance |
| /pack | /pack NVDA segment_mix |
Extract one of 12 certified valuation input packs |
| /metric | /metric NVDA gross_margin |
Query exact EDGAR XBRL concept recall |
| /cagr | /cagr NVDA revenue FY2022 FY2025 |
Zero-error compound annual growth rate calculation |
| /eval | /eval cagr NVDA revenue FY2022 FY2025 |
Deterministic zero-error algebraic evaluation |
| /peers | /peers semiconductors 5 |
XBRL metrics across dynamic SIC industry peers |
| /overview | /overview |
Overview of all certified FinSec valuation packs |
| /backtest | /backtest AAPL -3.0 |
5-Year SQL alpha backtest for morning dip reversals |
| /book | /book NVDA 10 |
Level-2 order book depth and liquidity squeeze detector |
| /logs | /logs 50.0 |
Audit HFT execution logs for slippage $>50\text{ms}$ |
| /ingest | /ingest TSMC lead times extended to 22 weeks |
Ingest breaking news into Calera ICX A4 Lattice |
| /quote | /quote supply_chain.semiconductors 0 |
Quote exact document slot with zero LLM inference |
| /stats | /stats |
Inspect ICX A4 lattice simplicial nodes & memory bytes |
| /contradictions | /contradictions |
Detect $Betti-2$ topological guidance voids |
| /portfolio | /portfolio |
Inspect current portfolio equity, cash, and sector weights |
| /rebalance | /rebalance Trim tech by 5% |
Stage systematic macro rebalance with HMAC tokens |
| /execute | /execute LITHOS-STG-XXXX TOKEN |
Execute staged order with human confirmation token |
| /model | /model claude-3-7-sonnet anthropic KEY |
Dynamically switch active LLM reasoning model |
| /toggle | /toggle llm or /toggle icx |
Dynamically toggle subsystems on or off |
| /status | /status |
Full subsystem diagnostics and latency metrics |
| /reset | /reset |
Clear session turns while preserving lattice facts |
| /help | /help |
Display command help matrix |
| /exit | /exit or /quit |
Disconnect from terminal |
🔌 Connecting to Claude Desktop & Cursor
Connect to Claude Desktop
Add Project Lithos to your claude_desktop_config.json:
{
"mcpServers": {
"project-lithos": {
"command": "python",
"args": [
"/ABSOLUTE/PATH/TO/PROJECT_LITHOS/server.py"
],
"env": {
"PYTHONUNBUFFERED": "1",
"LITHOS_BROKER_ENV": "paper",
"GEMINI_API_KEY": "YOUR_GEMINI_API_KEY",
"ICX_API_KEY": "YOUR_ICX_KEY",
"FINSEC_STANDBY_MODE": "true"
}
}
}
}
🧪 Testing
Project LITHOS includes comprehensive test suites covering command handlers, dynamic status coordination, valuation packs, and prompt injection defenses:
# Run all tests
pytest
# Run tests with verbose output
pytest -v
📜 License & Pilot Terms
Project LITHOS is licensed under the Calera Labs Source-Available Evaluation & Educational License.
- Free to Use: You are free to clone, study, run, test, and pilot Project LITHOS for educational, research, academic, and internal pilot evaluation purposes with your own BYOK AI models.
- License Restricted & No Commercial Copying: Commercial redistribution, resale, or copying proprietary integration harnesses for commercial SaaS products without express written permission from Calera Labs is strictly prohibited.
- See the full LICENSE file for complete details.
📚 Additional Documentation
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。