CRM

Treat the CRM as the agent's working record, not merely a database the agent can call

The research agent is deployed separately on eve and uses durable sessions, tools, skills, and schedules to continue CRM work over time. Tasks live in a Postgres-backed queue, with leases and FOR UPDATE SKIP LOCKED allowing multiple dispatchers to claim disjoint work safely. When an account or contact should be revisited, the agent schedules a new dueAt with a reason instead of relying on a blanket cron rule, so follow-up cadence belongs to each unit of work.

Sources: [1][2]

Write customer facts from observable evidence instead of model confidence

For identity and enrichment work, tools do not accept confidence scores; they report what was actually observed, such as a CRM signature block or a GitHub account identity. The evidence layer decides whether a fact is strong enough to write, while weaker or conflicting findings become suggestions for a representative. The research sandbox also uses deny-all network egress and is never given DATABASE_URL, reducing the paths through which mailbox or CRM content could leave via arbitrary shell commands while still letting the agent use local text-processing tools.

Sources: [1][2]

For sales and RevOps teams that want research and CRM maintenance to run continuously

Comp AI CRM fits teams that want to start with information they already own — mail, meetings, threads, and signature blocks — and use optional external research only when useful. It is particularly relevant when autonomous enrichment is desirable but guessed customer facts are not: the Agent tab exposes the work being done and the system keeps provenance and unresolved suggestions visible. Google or Microsoft identity and mailbox integrations make it a practical comparison point for teams trying to move from a CRM that humans constantly maintain toward one where agents perform much of the research and upkeep.

Sources: [1][2]

It is intentionally single-tenant and operationally more involved than a conventional CRUD CRM

The codebase deliberately has no organization model and is designed as a single-tenant CRM. Running it involves an agent deployment, Next.js frontend, NestJS API, Postgres, authentication, and — in the documented production stack — services such as Vercel Sandbox, AI Gateway, and Blob. That architecture offers durable autonomous work, but it also requires more deployment and operational understanding than a simple CRM application. Optional research providers also bring their own pricing and data-handling terms.

The repository's default release branch is the last tagged release intended for running the product, while main receives unreleased work. Deployments should therefore evaluate the release branch rather than assuming main is the stable target. The code is MIT-licensed, but the terms of Google, Microsoft, Vercel, model providers, and other connected services are separate from the repository license.

Sources: [1][3]