3 min read

The headless CRM for your agent

Source: http://x.com/i/article/2041093431028817920

I just crawled through macOS kernel hell so your AI agent could read every CRM record like a file.

crm.cli is an open-source, headless CRM built for agents. No UI. No dashboard. Just a CLI and a FUSE-mounted filesystem.

The problem

Every CRM assumes a human is clicking buttons. The most "AI-friendly" thing they've done is ship an MCP server with 15 half-documented tools.

Your agent doesn't need a createContact tool.

It needs a /contacts/acme-corp.md file it can open, read, and edit.

How it works

Mount your CRM as a directory.

Contacts become files. Deals become files. Edit a file, the CRM updates. Update the CRM, the file changes. Unleash Claude Code on it to do deep research or mass updates.

$ ls ~/crm/contacts/ acme-corp.md jane-doe.md $ cat ~/crm/contacts/jane-doe.md # Jane Doe - Company: Acme Corp - Email: [email protected] - Stage: qualified

There's an llm.txt at the mount root — so when your agent lands in the directory, it immediately knows the schema, conventions, and how to work with the data. Zero onboarding.

Or use the CLI — crm contact add, crm deal show, crm pipeline view. Every operation as a command with structured output.

Why files win

Point Claude Code or Codex at ~/crm and it just works. No auth tokens. No API docs. No tool schemas. The agent reads contacts like it reads any markdown file.

But unlike a folder of random markdown, crm.cli enforces structure — and this is where it earns its keep over a spreadsheet.

Data normalization.

Paste a phone number in any format — (212) 555-1234, +1-212-555-1234, 212.555.1234 — it's stored as E.164 and searchable by any variant. Drop a LinkedIn URL, it extracts the handle. Websites get normalized. Emails get validated. Your data is clean whether a human entered it or an agent did.

Fuzzy dedup and entity merging.

Add a contact that looks suspiciously like one you already have? crm.cli catches it. When duplicates slip through, crm merge combines two records — relinking every deal, activity, and reference to the surviving entity. No orphaned data.

Pipeline and reporting.

Deal stages, win/loss tracking, conversion rates, deal velocity — the stuff you actually need a CRM for.

crm report gives you pipeline health, stage-by-stage breakdown, and forecast in your terminal. Every stage transition is logged as an activity, so you get a full audit trail for free.

The CRM adds guardrails, then gets out of the way.

Under the hood

On Linux, crm mount uses FUSE to present the SQLite database as a userspace filesystem — fast, well-supported, and fully bidirectional.

macOS is trickier. Apple killed kernel extensions, which breaks traditional FUSE. So crm.cli ships a lightweight NFS server written in Rust that presents the same SQLite data as a network filesystem. Your Mac thinks it's talking to a file server. It's actually talking to your CRM.

Both platforms get the same interface.

crm mount ~/crm just works — you don't have to care which transport is underneath. Everything syncs back to a single SQLite database that never leaves your machine.

Try it

To have your agent use it, just point it to: https://dzhng.github.io/crm.cli/ Or if you want to install it the old fashion way:

bash

npm install -g @dzhng/crm.cli
# or: bun install -g @dzhng/crm.cli

# Or install the compiled binary
curl -fsSL https://raw.githubusercontent.com/dzhng/crm.cli/main/install.sh | sh

Github repo here: https://github.com/dzhng/crm.cli And if you want to run it in an easy to use personal cloud server with claude code / codex all preconfigured for you, try out

duet.so