trycompai/crm: The CRM Where the Agent Does the Work

A deep dive into a durable, evidence-first customer system that flips the usual CRM model. Humans define intent. Agents gather proof. The database just records what survived scrutiny.

9 min read • View on GitHub • More from trycompai

A human executive stands beside an almost blank form while a mechanical research agent sorts labeled evidence into a thick ledger book. The image shows the core inversion in this CRM: the human sets intent, the agent gathers proof, and the record only changes when evidence survives review. — In this CRM, the empty form is the point. The agent does the work, and the ledger keeps only what can be defended.

Key Takeaways

trycompai/crm flips CRM from a human entry system into an agent execution system.
The project treats evidence as the unit of truth and rejects confidence scores as a write path.
Durability matters here because the agent survives interrupts, leases work safely, and resumes without losing state.
The architecture makes the product feel safer than a typical AI CRM because humans set intent and ambiguous data routes to review.

The CRM Reversal

Traditional CRMs make humans do the typing and software do the storing. trycompai/crm reverses that contract: the agent researches, enriches, and updates, while the database acts like a ledger of verified facts. That is why this repo is interesting before you even look at the stack.

The project’s basic claim is blunt. A CRM should not reward whoever can fill forms fastest. It should reward the system that can accumulate evidence, keep provenance, and leave a clean trail for review.

The agent is not a feature of the CRM; the CRM is where the agent keeps its notes.

Lewis Carhart, Co-founder & CEO, Comp AI · trycompai/crm README

Why Confidence Is the Wrong Primitive

The repo rejects the familiar AI habit of returning a confidence score and calling it rigor. Instead, the agent reports observations such as identity matches, signature blocks, and public records. The system then decides whether that evidence is strong enough to write back.

That choice changes the product’s epistemology. The goal is not to make the model sound sure. The goal is to prevent plausible lies from becoming customer data.

A close-up proof desk shows index cards of different evidence types moving toward a narrow gate, with a red stamp for insufficient evidence and a green stamp for write to record. Some cards stop at the threshold while others pass into a ledger channel. — The system prices evidence, not confidence. Weak signals remain visible, but they do not mutate the record.

This diagram shows the product’s real loop: observe, store provenance, test the threshold, then write only when the evidence clears.

How the Durable Queue Keeps the Agent Alive

The technical proof lives in the task system. The repo’s queue pattern uses PostgreSQL leasing with FOR UPDATE SKIP LOCKED, which lets multiple workers claim due work without stepping on each other. If an agent crashes, the task is not lost. It is simply eligible again.

// lib/tasks.ts style queue claim pattern
const rows = await db.$queryRaw`
  SELECT id
  FROM tasks
  WHERE due_at <= now()
    AND leased_at IS NULL
  ORDER BY due_at ASC
  FOR UPDATE SKIP LOCKED
  LIMIT 1
`;

// lease task, process, then write evidence back when complete

That is a small detail with big consequences. It turns the agent into durable infrastructure instead of a fragile cron job with an LLM attached.

Pattern	What happens	Failure mode
Cron job	Runs on a timer and hopes work is still valid	Missed work or duplicate work
Leased queue	Claims one due task at a time	Work resumes safely after interruption
Request-response AI	Answers once and exits	State disappears with the request
Durable agent	Keeps state and returns to the queue	Slow work survives restarts

What Eve Changes About Agent Design

The other unusual move is filesystem-first agent design. Under .agents/, the repo stores SKILL.md files and AI elements as readable artifacts, which means behavior is versioned like code but read like prose. That is a much more legible contract than hidden prompt strings buried in application logic.

This matters because agent behavior is not just a model choice. It is a repository of instructions, tools, and conventions that teammates can inspect, review, and modify.

.agents/skills/nestjs-best-practices/SKILL.md is the kind of file that tells you the project wants agents to be teachable, not magical.

Why This Architecture Feels Safer Than a Typical AI CRM

The sandbox posture helps. The agent operates with deny-all egress, so the default is containment rather than open-ended network freedom. Add the no-guessing rule and human intent at the top of the workflow, and the whole system reads as controlled automation instead of autonomous improvisation.

Nothing about a person is guessed. No tool accepts a confidence score, because a model asked to grade its own certainty will, and it will be wrong in the direction that makes it look useful.

Lewis Carhart, Co-founder & CEO, Comp AI · trycompai/crm README

That is a strong governance stance for a CRM. It does not promise omniscience. It promises fewer silent mistakes.

How It Compares to Twenty, HubSpot, and Attio

This project is not trying to win the CRM market by being the most complete dashboard. It is trying to change who does the work. Most CRMs still optimize for human data entry. trycompai/crm optimizes for machine verification.

System	Primary worker	Data model	Workflow style	Self-hosted?	Evidence policy	Best fit
trycompai/crm	Agent	Evidence ledger	Durable research and review	Yes	No guessing, write only on proof	Teams that want autonomous enrichment
HubSpot	Human	Customer database	Manual plus automation	No	Confidence lives in features, not the core model	Broad sales and marketing ops
Twenty	Human	Flexible CRM schema	Traditional CRM workflow	Yes	User-driven data management	Open-source CRM teams
Attio	Human	Unified data model	AI-assisted workflow	No	Strong automation, still SaaS-first	Modern teams wanting polished UX

The contrast is not really about features. It is about labor economics and trust. Self-hosting, evidence thresholds, and agentic durability make this project feel closer to a controlled internal system than a generic SaaS product.

What the Project Suggests About the Next CRM Category

The larger implication is simple. CRM may stop being a place where people type account truth into forms. It may become a system that continuously assembles, verifies, and audits customer intelligence in the background.

The pitch is literal, not marketing fluff: a durable research agent is the product; the database is where it writes things down.

Yash Thakker, Tech Analyst · Comp AI Open-Sourced an Agentic CRM

That is the cleanest way to describe the repo. The database is not the product. The verified record is the product, and the agent is the worker that earns it.