odoo-feeder

AI-driven tool to populate an Odoo demo database with realistic, industry-specific data — built for sales, no technical setup required.

Launch (one line)

bash <(curl -fsSL https://raw.githubusercontent.com/odoo-ps/odoo-feeder/main/feed.sh)
# ...or, if you have wget instead of curl:
bash <(wget -qO- https://raw.githubusercontent.com/odoo-ps/odoo-feeder/main/feed.sh)

It installs what's missing (agy, bubblewrap, Node.js, Python, and optionally gum for nicer prompts), fetches the feeder, and runs it. You can pass the details as flags to skip the prompts:

bash <(curl -fsSL https://raw.githubusercontent.com/odoo-ps/odoo-feeder/main/feed.sh) \
  --url https://mycompany.odoo.com --login admin --secret <api-key> \
  --scope "Bakery" --company "Maison Rorive" --website https://www.maisonrorive.be

First run only: launch agy once and sign in with Google. The public odoo.com site is blocked as a target.

Templates

Instead of answering the questions one by one, you can describe the whole demo in a template — a free-text block the AI reads directly. Interactively, pick "Paste a template" and paste it; non-interactively, pass --template FILE.

name: Maison Rorive demo
website: https://www.maisonrorive.be
company_name: Maison Rorive
country: BE
apps: [sale_management, crm, website, helpdesk]
data_size: big
override_model_size: {product.product: 100, crm.lead: 30}
configuration: enable pricelists, multi-currency, default sales team "Retail"

Any key you write is understood (apps → modules to install, data_size → scale, override_model_size → exact per-model counts, country → company config, …), and free text is followed as instructions. Connection details (URL / login / secret) are always asked separately and never belong in a template.

Ready-made templates for common demo profiles live in templates/ (e.g. --template templates/research-institute-hr.tpl); copy one and adjust the name / website / company_name keys for the target prospect.

The prompts use gum for a nicer interface when it is available, and fall back to plain prompts otherwise.

AI CLI providers

By default the feeder drives agy (Antigravity). Pass --ai-cli copilot (GitHub Copilot CLI) or --ai-cli claude (Claude Code) to use a different one.

copilot can also be routed through OpenRouter instead of GitHub's own model routing, via Copilot's BYOK support. The API key is never passed as a flag or plaintext env var — store it once in the OS keyring:

keyring set odoo-feeder openrouter-api-key

then run with:

odoo-demo-feeder --ai-cli copilot --openrouter-model anthropic/claude-sonnet-4.5 ...

(OPENROUTER_MODEL env var works the same way.) No GitHub sign-in is required once OpenRouter is active.

Debugging a run

Headless runs only print whatever the model chooses to narrate — which can look like progress even when nothing actually happened (e.g. a weak model fabricating tool calls, or a real tool call that failed and got glossed over). Every run's full transcript + a ground-truth audit trail of every real odoo-crud call (and its exit code) is saved to ~/.local/share/odoo-demo-feeder/logs/; the feeder prints a file:// link to it at the end of every run, success or failure. If the summary claims success but no write actually succeeded, the feeder catches that itself and reports failure instead of "✔ Done!".

What's inside

• feed.sh — one-shot bootstrap (provision + launch).
• odoo-demo-feeder — the launcher: prompts/flags, refreshes the skill, runs the AI agent (headless by default, -i for interactive) inside a bubblewrap sandbox.
• odoo_crud.py — the only gateway the agent uses to reach Odoo (XML-RPC CRUD, method calls, CSV import, introspection).
• odoo-demo-csv/ — the npx skills package the agent follows to research a website and generate/import Odoo CSVs.

Safety

• The agent is restricted to a single tool (odoo-crud) via agy's permission allowlist, and confined by bubblewrap (read-only filesystem, secrets hidden, only the database credentials exposed).