Your team already lives in Odoo, but the intelligence sits outside it. People copy a customer's history into a separate chatbot, paste a draft back into a mail.message, or re-key numbers from a list view into a spreadsheet to get a summary. Odoo core ships no large-language-model layer at all — there is no place to ask "summarize this opportunity", "draft a reply to this ticket in our tone", or "explain why this invoice is overdue" without leaving the record. A typical off-the-shelf apps.odoo.com module bolts on a generic chat widget with one hard-coded provider and no awareness of your access rights, your custom fields, or your workflows, so it stalls the moment you need it to actually touch your data. This is where a made-to-order solution earns its keep.

ECOSIRE builds a native agentic AI assistant that runs inside the Odoo web client and speaks your database. At its core is a provider-agnostic connector layer: your API keys for the model families you choose (the ChatGPT, Claude, and Gemini families are all supported) are stored server-side via ir.config_parameter or a dedicated encrypted settings model, never shipped to the browser. On top of that sits a chat model (models.Model with threaded history on the mail.thread backbone) and an agent runtime that exposes a curated set of tools — each tool is a Python method that reads or writes specific models through the ORM, so the agent can fetch a partner's open orders, draft a QWeb-rendered quote summary, or post a follow-up on a lead. An OWL-based chat panel gives users a familiar conversational surface docked in the backend, with streaming responses and per-conversation history.

Technically, the module is a clean Odoo addon: a proper __manifest__.py declaring dependencies (base, mail, and whichever business apps your tools touch), models under models/ using fields, compute methods with @api.depends, and controllers exposing JSON-RPC endpoints that the OWL client calls. Security is first-class rather than an afterthought — the assistant runs tools in the session user's own environment, so ir.model.access.csv and record rules are enforced automatically: a salesperson's agent can never read HR payroll, because the ORM refuses it under that user's rights. We add explicit tool-level guardrails on top (an allowlist of callable methods, read-only vs write modes, confirmation prompts before any state-changing action, and full audit logging of every prompt, tool call, and token count). The design works on both Community and Enterprise; where Enterprise-only apps are in scope we wire against them, and where you are on Community we keep tools to the models you actually run.

Because this is built to order, delivery is a short engineering project, not a download. After a scoping call we confirm which providers, which models, and which tools (the exact ORM actions the agent may take) belong in your build, then develop against a staging copy of your database, run UAT with your team, and deploy with a rollback plan. Typical delivery is 2–4 weeks from confirmed scope. Pricing starts from $249 (indicative, single-company base scope with one provider and a small set of read-only tools); additional providers, write-enabled or cross-app agent tools, multi-company routing, and deeper data-access integrations increase the quoted scope. You receive a fixed quote after the scoping call, never a surprise invoice.