Comp AI CRM is a self-hostable customer relationship manager organized around a research agent rather than manual data entry. The agent runs as its own deployment on its own schedule and work queue: it decides which contacts and companies to look at next, researches them within a budget, records what it finds, and books its own follow-up checks. The CRM records are where the agent keeps its notes, and reps can watch or talk to it from an Agent tab on every contact, company, and deal.
Its central rule is that nothing about a person is guessed. Tools report what they observed (for example a signature block in an email or a GitHub account identity) rather than a confidence score, and a ledger weighs that evidence. Strong evidence writes to the record, and weak evidence becomes a suggestion for a human to accept or reject.
The agent is built on eve, Vercel's filesystem-first framework for durable agents, with 18 tools, four markdown skills, and a dispatcher that leases due tasks from a Postgres queue. It reads the team's own email threads, meetings, and signature blocks first, and each optional key adds a source: Perplexity for web search with citations, and Context for company brand data and LinkedIn profiles. Its sandbox has a shell but no network egress and no database credentials, so customer email content cannot leave through a shell command.
The rest of the stack is a Turborepo monorepo on Bun: a Next.js front end with shadcn/ui, a NestJS API over tRPC with mailbox sync, Prisma on Postgres, optional Redis, Better Auth for sign-in through Google, Microsoft, or a custom identity provider, and Vercel AI Gateway for model access. It is single-tenant by design and deploys to Vercel.
Features
- Autonomous research agent: runs on its own schedule against a work queue, spends a research budget, and keeps going after the browser closes
- Evidence ledger: tools report observations rather than confidence scores, strong evidence updates records and weak evidence becomes a human-reviewed suggestion
- Self-scheduled follow-ups: the agent books rechecks on contacts and shows the rep its stated reason
- Agent tab: every contact, company, and deal shows the agent's steps, discarded leads, and questions, and accepts durable conversations
- Inbox-first enrichment: reads the team's own threads, meetings, and signature blocks before any paid data source
- Optional sources: web search with citations through Perplexity, and company brand data plus LinkedIn profile lookups through Context
- Isolated sandbox: a shell with deny-all egress and no database credentials for working on dossiers and threads
- Typed stack: Next.js front end and NestJS API over tRPC, with Prisma on Postgres and Better Auth sign-in
