CRM by Comp AI

Comp AI's CRM is an open-source, MIT-licensed project that inverts the traditional CRM model: instead of a database with a form in front of it, the agent is the product and the database is simply where it writes things down. Built by the trycompai organization on GitHub, it has attracted over 3,300 stars since its launch in late July 2026, signaling strong early interest from the developer community.

What It Is

CRM is a self-hosted, single-tenant customer relationship management system built around a durable research agent rather than manual data entry. The agent runs on its own deployment, on its own schedule, and against its own work queue — it decides what to look at next, books its own follow-ups, spends a research budget, and continues running even when the browser is closed. The API layer deliberately contains no intelligence: it reports that something happened (a thread was ingested, a company was created, an attendee is unknown), and the agent decides what it means.

How the Agent Works

The agent is built on eve, Vercel's filesystem-first framework for durable agents, and ships with 18 authored tools, 4 skills (prose files the agent reads, versioned like code), one dispatch schedule, and a sandboxed shell environment.

Key design principles the README documents:

• Nothing about a person is guessed. Tools report only what they observed; a ledger prices the evidence. Strong evidence writes to the record; weak evidence becomes a human-settleable suggestion.
• Every outside source is optional. With no API keys, the agent still works by reading your own threads, meetings, and signature blocks. Each additional key (LinkedIn via RapidAPI, Perplexity for web search, Context for company brand data) opens one more research source.
• The sandbox has no network and no database. This removes the only path by which a customer's email body could leave through a shell command, and prevents a shell with credentials from becoming an exfiltration vector.
• The work queue is durable. claimDue uses FOR UPDATE SKIP LOCKED so two dispatchers take disjoint work and a crashed run frees its row when the lease expires.

Every contact, company, and deal has an Agent tab showing steps as they happen, leads discarded and why, and questions the agent cannot resolve surfaced for a human to settle.

The Stack

The project is a Turborepo monorepo on Bun, deployed on Vercel, with three independent deployments (Next.js front end, NestJS API, and the agent) sharing a Postgres database.

The front end communicates with the API over tRPC, with the router type generated from NestJS routers, making the stack type-safe from the Prisma row to the table cell.

Setup Path

Getting started requires Bun and Docker. After cloning and copying .env.example, only four values are required: BETTER_AUTH_SECRET, ALLOWED_SIGN_IN (the entire authorization model — an unset value means nobody can sign in), and a Google OAuth client ID/secret pair. DATABASE_URL defaults to the Docker Compose Postgres. Everything else — Perplexity, RapidAPI, Redis, the agent bridge — is optional and the app runs without any of it.

Current Status

The repository was created on July 31, 2026 and last pushed on August 3, 2026, making it a very recently launched project. It is licensed MIT, actively maintained, and open to contributions — the CONTRIBUTING.md explicitly notes a preference for human-written pull requests over agent-generated ones.