Operations teams want AI on their own data — an agent that drafts a quotation follow-up from the sale.order, triages an inbound helpdesk.ticket, or summarizes a vendor account.move — but every one of those ideas today means a developer ticket. Odoo Enterprise 19 ships an AI field assistant and OpenAI-backed helpers, but they are point features tied to Odoo's own model choices; there is no place for a non-developer to say "here is my agent, here is its prompt, here are the records it can read, here is which LLM runs it." Native Odoo runs out of road exactly where you need to compose behavior: automated actions and server actions can call Python, but authoring, versioning, and cost-governing an LLM agent through the UI is not something the base product does.

This module adds that missing configuration layer. We build a new model (ecosire.ai.agent) whose records ARE your agents: each holds a display name, a provider/model selection (selection field spanning OpenAI GPT, Google Gemini, and Anthropic Claude), a system prompt and response-style config, a temperature/token budget, and a set of Many2one/Many2many bindings that scope which Odoo models and records the agent may read. Agents are authored in a clean OWL/XML backend view — no code, no server restart. At run time the agent is invoked from a smart button, an ir.actions.server automated action, or the JSON-RPC/XML-RPC API; the module marshals the bound record's fields into the prompt, calls the selected provider over HTTPS, and writes the response back to a linked record, a chatter message, or a mail.activity. Every call is metered: a computed cost field (@api.depends on token counters) and a per-agent usage log give operations a real spend view instead of a surprise invoice.

Security is first-class because agents touch real business data. Access is governed by ir.model.access.csv for the agent and log models plus ir.rule record rules so, for example, a sales agent is only ever bindable to sale.* records the invoking user can already see — the agent never becomes a privilege-escalation path around Odoo's own ACLs. API keys are stored as encrypted ir.config_parameter values, never in plaintext views, and outbound calls are gated so a misconfigured agent cannot exfiltrate fields outside its declared binding. QWeb templates render the optional agent-run report for audit.

Because this is a build-to-order product, we scope it to YOUR agents and YOUR models before we write a line of code. You tell us the first two or three agents you actually need (which app, which trigger, which output), we confirm the __manifest__.py dependency list and target Odoo version (17.0, 18.0, or 19.0; Community or Enterprise), and ECOSIRE builds, tests on a staging copy of your database, and installs it. There is no instant download — typical delivery is one working week from confirmed scope — and you receive the full source so you are never locked to us.