The CRM That Works While You're Not Looking

You know the drill with CRMs. You spend more time maintaining the database than actually talking to customers, and the AI features are just a chat box bolted onto a form. The real work—figuring out what's true and writing it down—still lands on you. Comp AI CRM flips that around: it's an open source CRM where the agent isn't a feature, it's the thing doing the work.

What It Does

Comp AI CRM is an agentic-first CRM built for AI agents rather than humans. The agent runs on its own deployment, on its own schedule, against its own work queue. It decides what to look at next, books its own follow-ups, spends a research budget, and stops when the budget runs out. Close the browser and it keeps going—nothing about it is request-response.

The agent lives in apps/agent, a separate deployment built on eve, Vercel's filesystem-first framework for durable agents. A tool is a file, a skill is a markdown file, a schedule is a file, and the runtime handles the durable part: sessions that survive a redeploy and work that resumes where it stopped. The project ships with 18 authored tools (read_crm_history, search_crm, identify_contact, research_person, enrich_company, record_fact, schedule_recheck, and more), 4 skills written as prose markdown files (evidence.md, identity-matching.md, data-boundaries.md, writing-a-brief.md), and a single schedule file (dispatch.ts) that decides nothing—it just leases what's due and starts a session per row. The stack is Bun for the runtime and Postgres for the database, under an MIT license.

Why It's Cool

Nothing about a person is guessed. This is the rule the agent never breaks, and it's the most interesting design decision in the project. No tool accepts a confidence score, because—as the README puts it—a model asked to grade its own certainty will, and it will be wrong in the direction that makes it look useful. Instead, tools report what they observed (things like crm.signature-block or github.account-identity), and a ledger prices the evidence. Strong evidence writes to the record. Weak evidence becomes a suggestion a human settles. A confidently wrong fact about a customer is worse than a blank field, because nobody can tell it's wrong.

• The agent has a budget and a work queue. It's not sitting around waiting for you to ask it something. It leases due work, runs sessions, and stops when the research budget runs out. That's a fundamentally different model from the chat-box approach most "AI CRMs" take.

• Skills are versioned prose. The four skill files aren't code—they're markdown the agent reads, versioned like code. That means you can review, diff, and iterate on the agent's behavior the same way you'd review a pull request.

• Durable by default. Because it's built on eve, sessions survive redeploys and work resumes where it stopped. You don't have to babysit a long-running enrichment job.

• An evidence ledger instead of vibes. The separation between observation, pricing, and human-settled suggestions is a genuinely thoughtful architecture for the hardest problem in AI-assisted data entry: trust.

How to Try It

The README points to a quick start and configuration section, so the fastest path is to head to the repo and follow along there:

1. Clone the repository from github.com/trycompai/crm.
2. Check the README's Quick start section for the exact setup commands—the agent and the app are separate deployments, so expect to configure both.
3. Read the Configuration section for environment and database settings (you'll need Postgres).
4. When you're ready to run it for real, the Deploying section covers taking it live.

There's also a CONTRIBUTING.md if you want to get involved, and the agent's tools, skills, and schedule all live in apps/agent if you want to see how it's put together before you run anything.

Final Thoughts

This is a project for developers who've looked at the CRM category and thought the AI layer was backwards—and who care more about data you can trust than features that demo well. The evidence-ledger approach and the "agent keeps its own notes" framing are the kind of design choices that come from people who've actually been burned by confident, wrong data. It's early, and the README is honest about that, but if you're building anything where an agent needs to maintain a record of real people, the patterns here are worth studying—and running.

Follow @githubprojects for more developer tools and open source projects.