ProspectAI - Multi-Agent Investment Analysis System
🚀 Try It Now
Try the ProspectAI Web Demo — Experience the agentic system in action!
Overview
ProspectAI is a multi-agent investment analysis system built on the CrewAI framework. It leverages five specialized AI agents running a six-task pipeline to produce investment recommendations through a systematic analysis workflow. The system supports Anthropic Claude models (default) and local Ollama models.
Current release: v1.13.0
⚠️ Important Disclaimer
ProspectAI is built for educational purposes to help developers get initiated in Agentic AI development.
🚨 INVESTMENT WARNING: This tool should NOT be used as an investment tool without proper knowledge of the investment domain. The analysis provided is for educational demonstration of AI capabilities and should not be considered as financial advice. Always consult with qualified financial professionals before making investment decisions.
Features
- Multi-Agent System: Five specialized AI agents across a six-task pipeline for rigorous investment analysis
- Anthropic Claude: Powered by Claude models (Haiku 5.5 for the market, technical and fundamental steps, Sonnet 5.5 for the draft, critic and final steps; other Sonnet and Opus models supported)
- Ollama Support: Run fully locally with Ollama models
- Serper Web Search Discovery: Candidate stocks come from three fixed Serper web searches per run
- News Sentiment: Each picked stock gets the Marketaux news sentiment score for the last 7 days (optional
NEWS_SENTIMENT_API_KEY; null when absent or when too few articles exist) - Sector Analysis: Analyze 5 major sectors (Technology, Healthcare, Finance, Energy, Consumer)
- Command-Line Interface: Easy-to-use CLI with flexible configuration
- Environment-Based Config: Secure
.env-based configuration with startup validation - Structured Output: Consistent, machine-readable JSON analysis results
- CrewAI Framework: Professional multi-agent orchestration via LiteLLM
Architecture
The system runs five specialized agents across a six-task sequential pipeline. Each task receives the full output of all prior tasks as context.
MarketAnalyst → TechnicalAnalyst → FundamentalAnalyst → DraftStrategist → Critic → FinalStrategist
Task 1 Task 2 Task 3 Task 4 Task 5 Task 6
Pipeline Tasks & Agents
| Task | Agent | Tools | Input context |
|---|---|---|---|
| 1. Market Analysis | MarketAnalyst | PatientSerperDevTool |
— |
| 2. Technical Analysis | TechnicalAnalyst | TechnicalAnalysisTool, TechnicalInterpretationTool |
Task 1 |
| 3. Fundamental Analysis | FundamentalAnalyst | FundamentalDataTool, FundamentalGraderTool |
Tasks 1–2 |
| 4. Draft Strategy | InvestorStrategic | CompositeScoreTool, PortfolioAllocatorTool |
Tasks 1–3 |
| 5. Critic Review | Critic | — (synthesis only) | Tasks 1–4 |
| 6. Final Strategy | InvestorStrategic | CompositeScoreTool, PortfolioAllocatorTool |
Tasks 1–5 |
Agent Descriptions
Market Analyst Agent (Task 1)
- Purpose: Entry point — identifies candidate stocks from live market signals
- Function: The Flow searches the web and the model picks trending stocks from the results, incorporating macro/geopolitical context at execution time
- Data Sources: Serper web search (three fixed queries)
- Sentiment: After the model chooses its picks, the pipeline looks up each pick's news sentiment from Marketaux (recent articles about the company, cached per symbol per UTC day). The model never writes a score. A stock with no articles in the window, or no lookup, has a null sentiment and the run continues
- Output: Top 5 candidate stocks with sentiment scores (null when unavailable) and relevance metrics
Technical Analyst Agent (Task 2)
- Purpose: Quantitative technical analysis
- Function: Runs 13+ technical indicators per ticker using
yfinance+talibrary, then interprets signals usingTechnicalInterpretationTool - Indicators: RSI, MACD, Bollinger Bands, ATR, SMA, EMA, VWAP, ADX, and more
- Output: Per-stock signals, momentum scores (1–10), market regime (TRENDING/REVERTING), entry zones, stop-loss levels
Fundamental Analyst Agent (Task 3)
- Purpose: Financial statement and valuation analysis
- Function: Fetches real P/E, margins, debt ratios, FCF, and growth rates via
yfinance; grades each metric usingFundamentalGraderTool - Output: Valuation grades (CHEAP/FAIR/EXPENSIVE), financial health ratings, growth outlook
Investor Strategic Agent — Draft (Task 4)
- Purpose: First-pass portfolio construction
- Function: Computes composite scores via
CompositeScoreTool, allocates positions viaPortfolioAllocatorTool - Composite Score: 30 pts sentiment + 40 pts momentum + 30 pts fundamentals (max 100)
- Recommendations:
STRONG_BUY / BUY / HOLD / REDUCE / AVOID— allocation cap 40% per position - Trade setups:
LONG-BUY,SCALED-ENTRY,WAIT-FOR-ENTRY
Critic Agent (Task 5)
- Purpose: Adversarial quality gate between the draft and final strategy
- Function: Reviews the draft portfolio against 12 failure-mode checks (schema validity, allocation math, entry zone logic, R/R ratios, regime consistency, etc.) and emits structured directives for correction
- Output:
CriticOutput— list ofCritiqueItemobjects withfield,severity, anddirective;approvedflag if no issues
Investor Strategic Agent — Final (Task 6)
- Purpose: Revised, production-ready portfolio
- Function: Incorporates all Critic directives and produces the definitive
InvestorStrategicOutput— identical schema to the draft but corrected - Output: Machine-readable JSON (
pipeline_version: "2.0") with allocation percentages summing to 100%, post-generation validation warnings attached
Installation
Prerequisites
- Python 3.9+
- An Anthropic API key or a local Ollama installation
Install via pip
pip install prospectai
That's it. All dependencies (CrewAI, yfinance, ta, requests, etc.) are installed automatically.
Configure Environment
Create a .env file in the directory where you'll run prospectai (see env.example for the full template):
# LLM backend: anthropic | ollama
MODEL_PROVIDER=anthropic
# Optional on Anthropic: a global model id for all agents (raw id, no provider prefix).
# Unset, each agent uses its model from config/agents.yaml (claude-haiku-5-5 for market,
# technical and fundamental; claude-sonnet-5-5 for investor_strategic and critic).
# MODEL=claude-sonnet-4-6
# Required when MODEL_PROVIDER=anthropic
ANTHROPIC_API_KEY=your_key_here
# Optional: override MODEL for a specific agent only
# AGENT_MARKET_ANALYST_MODEL=claude-sonnet-4-6
# AGENT_TECHNICAL_ANALYST_MODEL=
# AGENT_FUNDAMENTAL_ANALYST_MODEL=
# AGENT_INVESTOR_STRATEGIC_MODEL=
# Market data — Serper web search (required)
SERPER_API_KEY=your_key_here
# When MODEL_PROVIDER=ollama
# OLLAMA_BASE_URL=http://localhost:11434
Model resolution (simple rule): AGENT_*_MODEL (if set for that agent) → MODEL → legacy ANTHROPIC_MODEL / OLLAMA_MODEL if you still use older .env files. MODEL_PROVIDER selects Anthropic vs Ollama routing.
When you run the CLI, MODEL is also set to provider/model_id for CrewAI’s environment fallback; the app still resolves raw ids for each agent from the variables above.
Getting your API keys
- Anthropic: console.anthropic.com
- Serper (optional fallback): serper.dev
Ollama setup (optional — local models)
# Install Ollama
# https://ollama.com/download
ollama serve
ollama pull qwen3.5:9b
Usage
# Analyze Technology sector (default)
prospectai
# Analyze a specific sector
prospectai --sector Healthcare
prospectai --sector Finance
prospectai --sector Energy
prospectai --sector Consumer
# Override the global MODEL for a single run
prospectai --model claude-opus-4-6 --sector Technology
# Use a local Ollama model
prospectai --ollama --sector Technology
prospectai --ollama --model llama3.2:8b --sector Healthcare
prospectai --ollama --url http://192.168.1.100:11434 --sector Finance
CLI Reference
| Flag | Description |
|---|---|
--sector |
Sector to analyze: Technology, Healthcare, Finance, Energy, Consumer (default: Technology) |
--model |
Override global MODEL for this run (raw model id) |
--ollama |
Use local Ollama instead of Anthropic |
--url |
Ollama server URL (overrides OLLAMA_BASE_URL) |
Configuration
LLM environment variables
| Variable | Required | Description |
|---|---|---|
MODEL_PROVIDER |
Yes | anthropic or ollama (CLI --ollama sets ollama) |
MODEL |
Ollama only | Optional on Anthropic (each agent defaults to its config/agents.yaml model: claude-haiku-5-5 for market, technical and fundamental, claude-sonnet-5-5 for investor_strategic and critic); a global override for all agents when set. Required with --ollama (e.g. qwen3.5:9b) |
ANTHROPIC_API_KEY |
If MODEL_PROVIDER=anthropic |
Anthropic API key |
OLLAMA_BASE_URL |
If MODEL_PROVIDER=ollama |
Ollama server URL |
AGENT_MARKET_ANALYST_MODEL |
No | Overrides MODEL for the Market Analyst agent only |
AGENT_TECHNICAL_ANALYST_MODEL |
No | Overrides MODEL for the Technical Analyst agent only |
AGENT_FUNDAMENTAL_ANALYST_MODEL |
No | Overrides MODEL for the Fundamental Analyst agent only |
AGENT_INVESTOR_STRATEGIC_MODEL |
No | Overrides MODEL for the Investor Strategic agent only |
ANTHROPIC_MODEL |
No | Legacy fallback if MODEL is unset (Anthropic path) |
OLLAMA_MODEL |
No | Legacy fallback if MODEL is unset (Ollama path) |
*For Ollama, set MODEL or the legacy OLLAMA_MODEL.
Other environment variables
| Variable | Required | Description |
|---|---|---|
SERPER_API_KEY |
Yes | Serper web search, the only discovery source for the market phase |
NEWS_SENTIMENT_API_KEY |
Optional | Per-stock news sentiment (Marketaux); without it sentiment is null. The free tier allows 100 requests a day shared by everyone using the key: when a day's quota runs out, the report shows null sentiment and execution_metrics.sentiment_lookup.stopped_reason says why. A stock needs at least one article from the last 7 days to get a score |
Anthropic Claude Models
| Model | Use Case |
|---|---|
claude-haiku-5-5 |
Default for the market, technical and fundamental steps (set in config/agents.yaml) |
claude-sonnet-5-5 |
Default for the draft, critic and final steps (set in config/agents.yaml) |
claude-sonnet-4-6 |
Higher quality, slower and costlier |
claude-opus-4-6 |
Highest quality, deeper reasoning |
Ollama Models
| Model | Notes |
|---|---|
qwen3.5:9b |
Good reasoning, recommended for analysis |
llama3.2:8b |
General purpose |
llama3.2:3b |
Lightweight, fast |
mistral:7b |
Good for analytical tasks |
Per-agent model overrides
Preferred: set optional AGENT_*_MODEL variables in .env or your host secrets (e.g. Modal). Each variable overrides the global MODEL for that agent only.
Fallback: you can still set llm.model per agent in config/agents.yaml if no env-based id is resolved for that agent.
See AGENT_LLM_CONFIGURATION.md for additional notes on agent LLM configuration.
Development
To contribute or run from source:
git clone https://github.com/moisesprat/ProspectAI.git
cd ProspectAI
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -e .
After installing in editable mode the prospectai command is available and picks up any local changes immediately.
Project Structure
ProspectAI/
├── agents/ # AI agent implementations
│ ├── base_agent.py
│ ├── market_analyst_agent.py
│ ├── technical_analyst_agent.py
│ ├── fundamental_analyst_agent.py
│ ├── investor_strategic_agent.py
│ └── critic_agent.py # Adversarial reviewer (v1.5.0+)
├── config/
│ ├── agents.yaml # Agent behavior (role, goal, model, thinking)
│ ├── tasks.yaml # Task definitions (descriptions, output schemas)
│ ├── agent_config_loader.py
│ ├── task_config_loader.py
│ └── config.py
├── schemas/
│ └── agent_outputs.py # Pydantic output schemas for all 5 agents
├── utils/
│ ├── technical_analysis_tool.py # 13+ indicator calculator
│ ├── technical_interpretation_tool.py
│ ├── fundamental_data_tool.py # yfinance financial data fetcher
│ ├── fundamental_grader_tool.py # Grades individual fundamental metrics
│ ├── composite_score_tool.py # 30/40/30 composite scoring
│ ├── portfolio_allocator_tool.py # Allocation math with 40% cap
│ └── recommendation_validator.py # Post-generation validation
├── scripts/
│ └── deploy.sh # Bump version, build, publish to PyPI, deploy Modal
├── tests/
├── main.py # CLI entry point (prospectai command)
├── prospect_ai_crew.py # CrewAI orchestration (6-task pipeline)
├── pyproject.toml
└── README.md
Running Tests
python tests/test_skeleton.py
python tests/test_technical_analyst.py
python tests/test_market_analyst_llm.py
Building and Publishing
pip install build twine
python -m build
twine upload dist/*
Troubleshooting
| Issue | Fix |
|---|---|
prospectai: command not found |
Run pip install prospectai or activate the venv where it's installed |
.env file not found |
Create a .env file in your current working directory |
| Missing keys error at startup | Check the listed keys against your .env (especially MODEL / MODEL_PROVIDER and market data keys) |
ANTHROPIC_API_KEY invalid |
Verify key at console.anthropic.com |
| Ollama connection refused | Run ollama serve and verify OLLAMA_BASE_URL |
| Ollama model not found | Run ollama pull <model-name> |
| No stocks found from search | Check SERPER_API_KEY and the Serper quota |
Sentiment is always empty (null, sentiment_available false) |
Check that NEWS_SENTIMENT_API_KEY is set and valid, then read execution_metrics.sentiment_lookup: status no_key means no key, stopped_reason quota or rate_limited means the provider's daily free-tier quota ran out (it resets daily; results already looked up that UTC day are cached), access_denied means the key was refused, and scores below the number of picks means those stocks had no article in the last 7 days |
Release Notes
v1.13.0 — News sentiment for each picked stock
- Each stock the analysis picks now gets a news sentiment score from Marketaux, based on recent articles about the company, and the report shows it next to the stock.
- A stock with too few recent articles, or a day when the news source is unavailable or out of free requests, shows no score instead of an invented one, and the analysis still finishes.
- Stock picks come from web search alone, and the old forum source that had stopped returning data is gone, so runs no longer wait on it.
v1.12.3 — Analyses that run at the same time stay separate
- When several analyses run at the same moment, each now keeps its own price and company data, so one analysis starting can no longer discard data another is still using.
- Within a single analysis nothing changes: each stock's data is still fetched once and reused across the steps.
v1.12.2 — More accurate run figures and stricter stock picks
- The token and cost figures shown for a run no longer count the draft step a second time under the final step, so the per-step, per-model and total figures now match what each step really used.
- A sector fund or a market-wide fund picked as a stock is now set aside before the picks are checked, so it can no longer leave a run without stocks or lower the other picks' relevance.
- Common capitalised words and abbreviations such as AI, CEO and EV in web search results no longer count as stock mentions unless they are written as a ticker, for example $AI or NYSE: AI.
- An analysis for a sector the app does not support is now refused straight away instead of starting paid web searches, and the stored record of a run's web searches never keeps a web address.
v1.12.1 — Runs skip prompt caching
- Analyses no longer use prompt caching. Every step of a run sends a different prompt, so a cached prompt was never reused and only added cost; each run now costs a little less, and reports are unchanged.
v1.12.0 — A stronger model for the analysis, review and final steps
- The draft, review and final steps of every analysis now run on Claude Sonnet 5.5, a stronger model than before. The research steps that gather market, technical and fundamental data stay on the faster Claude Haiku 5.5.
- Reports are still built the same way: every number is computed from live data, and the model only writes the commentary and chooses each stock's action.
Roadmap
Next — Multi-sector analysis
- Support analyzing up to 3 sectors in a single pipeline run
- Stocks from all sectors compete on composite score for portfolio slots
- Cross-sector portfolio diversification rules
Future
- Integration with financial news APIs (Bloomberg, Reuters)
- Monte Carlo simulations for portfolio scenarios
- Advanced risk metrics (VaR, CVaR, Sharpe ratios)
- PDF report generation with charts
Contributing
- Fork the repository
- Create a feature branch
- Make your changes with tests where applicable
- Submit a pull request
License
MIT License — see LICENSE for details.
Acknowledgments
- Built on the CrewAI framework
- LLM routing via LiteLLM
- Inspired by modern multi-agent AI systems
- [Author webpage](https://moisesprat.dev
Metadata
Release files for prospectai 1.13.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| prospectai-1.13.0.tar.gz | 272.0 kB | Without provenance | Details for prospectai-1.13.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| prospectai-1.13.0-py3-none-any.whl | Python 3 | none | any | Details for prospectai-1.13.0-py3-none-any.whl |
Total release size: 433.2 kB
Release files / prospectai-1.13.0.tar.gz
| Download URL | prospectai-1.13.0.tar.gz |
|---|---|
| Size | 272.0 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
51510a3ec10c8f387168d9b60751cedbf89d25e98c7acb25f94ae98cdba93726
|
|
BLAKE2b-256 checksum How to use checksums |
2e3ce85947f4da89d84cae095dc11402feabf930bc8506dada7594110a4df131
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.11.17
|
Release files / prospectai-1.13.0-py3-none-any.whl
| Download URL | prospectai-1.13.0-py3-none-any.whl |
|---|---|
| Size | 161.3 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
330aa3563b0dd654bea079dd26924156bdc2fec538948e3de750299a1c5bce83
|
|
BLAKE2b-256 checksum How to use checksums |
916ea01cae745623369d30ac121757c5fbb624619e269631eb2060b18da31bb3
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.11.17
|
Security for prospectai 1.13.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.