agenthinkmesh/vc-performance-agent

VC Performance Agent

VC Performance Agent is a full-stack web application for venture capital decision support. It combines a deal-flow CRM with an LLM-backed "AI Copilot". The Copilot analyzes deals for behavioral biases and red flags, suggests due-diligence questions, and scores founders, alongside a research-insights library, a Monte Carlo portfolio simulator and fund performance metrics (IRR, MOIC, batting average). The target users are VC fund managers, investment teams and fund operations staff.

This is an application codebase, not a trained model. LLM calls go to an external OpenAI-compatible chat-completions endpoint.

Author: AgenThink, Kuwait City

The application lives in the nested folder vc-performance-agent/. Its own README is here: vc-performance-agent/README.md.

---

Key features

Features as implemented in server/routers.ts and client/src/pages/:

Area	What it does
Deal flow CRM	Deals move through a pipeline with statuses Sourced → Meeting → Due Diligence → Term Sheet → Closed / Passed (as coded in the schema and UI). Also covers companies, contacts, an activity timeline and pipeline statistics.
AI Copilot	copilot.analyzeDeal flags 7 bias types: gambling mentality, ego-driven decisions, network homogeneity, confirmation bias, anchoring, herd mentality, and over-reliance on gut feel. It also flags red flags. copilot.suggestQuestions generates due-diligence questions and copilot.detectPatterns analyzes patterns across the portfolio.
Deal evaluation	deals.analyzeWithAI scores deals from 0 to 100 using weighted criteria from the research notes: Team Quality 95%, Business Model 74%, Market Size 68%.
Founder assessment	Founder profiles with an LLM-generated 0–100 assessment (founders.assessWithAI) covering leadership, domain expertise, track record, team building, execution ability, and vision.
Research library	Curated research insights (power law, J-curve, founder assessment, behavioral biases), filterable by category.
Portfolio simulator	A server-side Monte Carlo simulation of portfolio outcomes from success rate and return-multiple parameters.
Performance dashboard	Fund metrics (IRR, MOIC, batting average) with history.
Due diligence	Per-deal checklists.

---

Architecture

• Frontend: React 19, Tailwind CSS 4, shadcn/ui, wouter routing, tRPC client
• Backend: Express with tRPC. The routers are auth, companies, deals, founders, contacts, activities, portfolio, dueDiligence, research, simulations, performance and copilot.
• Database: MySQL/TiDB via Drizzle ORM. Tables: users, companies, deals, founders, contacts, activities, portfolio_investments, due_diligence_checklists, research_insights, portfolio_simulations, fund_metrics.
• LLM: server/_core/llm.ts calls an OpenAI-compatible /v1/chat/completions endpoint with model gemini-2.5-flash.
• Auth: OAuth via the Manus platform (session cookie with JWT)
• Tests: Vitest (server/*.test.ts)

Repository structure

Path	Description
VC_PERFORMANCE_AGENT_EXECUTIVE_SUMMARY.md	Stakeholder overview: features, target users, methodology, roadmap
vc_research_findings.md	Research notes on VC returns (power law, the math of VC returns, VC decision-making surveys)
vc_success_factors.md	An analysis of the key drivers of venture capital success
vc-performance-agent/client/	React frontend. src/pages/ has Home, Deals, DealDetail, Companies, CompanyDetail, Founders, Portfolio, Performance, Research and Simulator.
vc-performance-agent/server/	Express + tRPC backend: routers.ts, db.ts, storage.ts, _core/ (LLM, OAuth, env, Vite integration)
vc-performance-agent/drizzle/	Schema, relations and SQL migrations
vc-performance-agent/shared/	Types and constants shared by client and server
vc-performance-agent/todo.md	Feature checklist (73 items done, 32 open)

---

Getting started

Requirements: Node.js 22+, pnpm, and a MySQL-compatible database.

cd vc-performance-agent
pnpm install
# create .env with the variables below (the upstream README refers to a .env.example that is not included)
pnpm db:push        # drizzle-kit generate && drizzle-kit migrate
pnpm dev            # development server

Other scripts: pnpm build, pnpm start (production), pnpm test (Vitest), pnpm check (type-check), pnpm format.

Environment variables read by server/_core/env.ts (names only): DATABASE_URL, JWT_SECRET, VITE_APP_ID, OAUTH_SERVER_URL, OWNER_OPEN_ID, BUILT_IN_FORGE_API_URL, BUILT_IN_FORGE_API_KEY.

---

Results / metrics

The files report no evaluation results for the AI Copilot or the scoring model. The criterion weights (95% / 74% / 68%) come from the research notes. The "80% of VC investments fail" framing is the project's own motivation statement, not a measured outcome of this tool.

---

Limitations

• Tied to the Manus platform. Authentication and the default LLM endpoint depend on Manus OAuth and the Manus "Forge" API, so running elsewhere requires replacing server/_core/oauth.ts and llm.ts, or pointing them at compatible services.
• Referenced files are missing. The upstream README mentions a .env.example, a seed-research.mjs seed script, a Dockerfile, and several specification documents (technical spec, dev guide, process flows, flowcharts). None of these are in this repository.
• Simplified simulator. The Monte Carlo simulation is a basic random-draw model and is not calibrated to real fund data.
• Unvalidated scores. Bias detection and scoring are LLM-generated judgments, not validated predictors. Treat them as prompts for discussion, not investment advice.
• Unfinished items. Several items in todo.md remain open, among them network analysis, reference-check tracking and industry benchmarking.

---

License

MIT (see vc-performance-agent/LICENSE).

Citation

@software{agenthink2026vcperformance,
  title        = {VC Performance Agent: AI-Assisted Venture Capital Decision Support},
  author       = {{AgenThink}},
  year         = {2026},
  address      = {Kuwait City, Kuwait},
  url          = {https://huggingface.co/agenthinkmesh/vc-performance-agent}
}