Vercel-deployed, Bun-powered CRM prioritizes an autonomous agent over traditional databases. The agent operates independently, managing research, scheduling, and data updates without human intervention. Unlike standard CRMs, it decides what to investigate next, allocates a research budget, and stops autonomously when funds deplete. The system uses a sandbox with no network or database access, treating external tools as optional. Evidence is strictly fact-based; no confidence scores or guesses are allowed. Tools report observations (e.g., `crm.signature-block`), and a ledger prices evidence strength. Weak data becomes human-settled suggestions. Security is strict: single-tenant, Google-only sign-in, and no external API keys required.

The agent’s workflow is code-driven. Tools like `read_crm_history` and `enrich_company` are markdown files, while schedules and tasks are file-based. The Nest JS API merely logs events (e.g., thread ingested) to a queue; the agent processes these. A `dispatch.ts` file manages work queues using database leases to prevent overlaps. The frontend is Next.js with URL-based state, ensuring views are shareable links. The agent’s tab tracks its decision process, preserving conversations via signed tokens.

Built on Vercel AI Gateway and Prisma/Postgres, the stack emphasizes durability. Sessions survive redeploys, and tools are simple files. The sandbox enforces `deny-all` egress, isolating the agent from external risks. Configuration is minimal—four environment variables control access. The agent plans work based on available tools (e.g., LinkedIn API via RAPIDAPI_KEY), printing available integrations at startup. Everything runs locally or in Vercel’s sandbox, with no external dependencies beyond optional keys.