avansaber/erpclaw

The AI-native ERP. Talk to your books in plain English.
They answer, they act, and they say no when the numbers do not add up.

See it work · The books say no · Quick start · Trusted with money · erpclaw.ai

You do not learn screens. You say what happened, and ERPClaw does the bookkeeping: it reads the ledger to answer questions. In the payment shown below, it asks before posting to the books.

These are real conversations from a live run on 9 October 2026 against a test company, Northwind Supplies. The animation shortens the output; the numbers are exactly what the books returned.

"Northfield's invoice NF-7781 bills all 300 HX-200 cartridges on PO-2026-00001 at $42.00. Post it."

The assistant asked, got a yes, and tried to post. The books refused:

Invoice qty exceeds received qty for item 'HX-200 Cartridge'.
Ordered: 300.00, Received: 0, Already invoiced: 0, Current invoice: 300.00

That is a three-way match on a bill linked to its purchase order. What you ordered, what arrived, and what the supplier billed have to agree before that bill reaches the books. Nothing arrived, so $12,600.00 stayed out of the ledger. The check is not an instruction in a prompt that a model can be argued out of. It is a rule inside the ledger, and the AI cannot talk its way past it. The assistant did the useful thing instead and offered the next step: record the goods when they arrive, then post the bill.

"Show me everything recorded in the last fifteen minutes."

02:39:56  create-purchase-invoice   draft bill from PO-2026-00001, $12,600.00   never posted
02:39:09  submit-payment            PAY-2026-00001, 3 ledger lines
02:38:55  update-payment            PAY-2026-00001, amount and allocation set
02:38:20  add-payment               PAY-2026-00001, $2,450.00 received
02:37:00  submit-purchase-order     PO-2026-00001
02:37:00  submit-sales-invoice      INV-2026-00001
          ...and the setup entries before them

Each of these changes left a timestamped entry that says what it was and which record it touched. The refused bill is right there too, as a draft that never reached the ledger.

About five minutes. ERPClaw is one product: install it once, then describe your business and it sets itself up.

1. Install (OpenClaw)

clawhub install erpclaw

This installs the foundation and creates the database on your own machine.

2. Tell it about your business

"I'm opening a retail store called Sunrise Goods in Portland, Oregon. Set me up."
→ creates your company, a US GAAP chart of accounts, fiscal year, and tax rates

3. Ask for your industry. Industry editions arrive the same way, with no second install:

"I'm a school"           → pulls the education edition
"Set me up for a clinic" → pulls clinical practice management

hermes skills tap add avansaber/hermes-skills
hermes skills install avansaber/hermes-skills/skills/erpclaw --force
export ERPCLAW_HOME=~/.hermes/erpclaw-home    # blank = ~/.openclaw/erpclaw
python3 ~/.hermes/skills/erpclaw/scripts/erpclaw-setup/db_query.py --action initialize-database

git clone https://github.com/avansaber/erpclaw ~/.hermes/skills/erp/erpclaw
export ERPCLAW_HOME=~/.hermes/erpclaw-home
python3 ~/.hermes/skills/erp/erpclaw/scripts/erpclaw-setup/db_query.py --action initialize-database

ERPClaw also works through any MCP client.

Everything posts to one shared set of books under the same rules.

• Posted amounts can never be edited. Cancelling posts a reversing entry and marks the original as cancelled, so the history stays whole.
• Money is exact. Amounts are exact decimals, never floating point, so cents never drift.
• The rules live in the ledger. Balanced entries and the three-way match are checked by the books themselves, whichever agent is asking.
• Every release is cryptographically verified.
• Your books stay with you. Self-hosted: the database lives on your own machine or server, and conversations go only to the AI model you choose.
• Open source, GPL v3. $0 forever. No tiers, no per-seat bill.
• Your database, your agent. SQLite by default, PostgreSQL fully supported. Works through OpenClaw, Hermes, or any MCP client.

flowchart LR
    A["You, in plain English"] --> B["Your agent<br/>OpenClaw, Hermes or any MCP client"]
    B --> C["ERPClaw action<br/>submit-payment"]
    C --> D{"You confirm<br/>Continue? y/N"}
    D -- yes --> E{"Validation rules<br/>balanced entry<br/>exact decimals"}
    E -- pass --> F[("Ledger<br/>amounts never edited")]
    E -- fail --> G["Refused<br/>nothing posted"]
    F --> H["Audit trail"]

Most business software was built for forms and screens, and the AI arrives later as a chat window on top. ERPClaw was built for agents from the first commit: the conversation is the front door, and the accounting rules are code that every request has to pass through.

Atrium, the ERPClaw app for Mac. Coming soon to the Mac App Store.

• Website: erpclaw.ai
• Docs: erpclaw.ai/docs
• Stripe Marketplace: marketplace.stripe.com/apps/erpclaw-accounting
• All repositories: github.com/avansaber
• OpenClaw: openclaw.org

GNU General Public License v3. Copyright © 2026 AvanSaber Inc. See LICENSE.txt.

_{Built by AvanSaber Inc.}